tobias/ForAug
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

arxiv V1

Browse Source
This commit is contained in:
Tobias Christian Nauen
2026-02-24 11:52:26 +01:00
parent 5c08f9d31a
commit 7e66c96a60
434 changed files with 6790 additions and 68375 deletions
Download Patch File Download Diff File Expand all files Collapse all files

856
sec/experiments.tex
View File

@@ -1,525 +1,417 @@
% !TeX root = ../main.tex
\begin{figure}[t]
\begin{minipage}[t]{.62\textwidth}
\captionof{table}{ImageNet results when training ViTs with different data augmentation pipelines.
\schemename consistently improves performance in low- and mid-augmentation regimes and remains complementary to strong augmentation pipelines, with larger gains for larger models.
}
\label{tab:imagenet-pipelines}
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccc}
\toprule
\multirow{2.5}{*}{Augmentation} & \multirow{2.5}{*}{MixUp} & \multirow{2.5}{*}{CutMix} & \multicolumn{3}{c}{Accuracy [\%] using} \\
\cmidrule(l){4-6}
& & & ViT-S & ViT-B & ViT-L \\
\midrule
Basic & \xmark & \xmark & $71.9 \pm 0.1$ & $69.5 \pm 0.2$ & $68.3 \pm 0.4$ \\
Basic + \schemename & \xmark & \xmark & $75.7 \pm 0.2$ & $75.5 \pm 0.6$ & $73.1 \pm 1.7$ \\
& & & \grntxt{$+3.8$} & \grntxt{$+6.0$} & \grntxt{$+4.8$} \\
\midrule
RandAugment & \xmark & \xmark & $76.3 \pm 0.5$ & $75.5 \pm 0.2$ & $74.7 \pm 0.4$ \\
RandAugment + \schemename & \xmark & \xmark & $78.0 \pm 0.1$ & $77.8 \pm 0.1$ & $78.0 \pm 0.6$ \\
& & & \grntxt{$+1.7$} & \grntxt{$+2.3$} & \grntxt{$+3.3$} \\
\midrule
Basic & \cmark & \cmark & $79.8 \pm 0.3$ & $78.6 \pm 0.4$ & $78.1 \pm 1.6$ \\
Basic + \schemename & \cmark & \cmark & $79.8 \pm 0.3$ & $81.6 \pm 0.5$ & $81.0 \pm 0.4$ \\
& & & \gtxt{$\pm 0.0$} & \grntxt{$+3.0$} & \grntxt{$+2.9$} \\
\midrule
3-Augment & \xmark & \cmark & $79.1 \pm 0.1$ & $77.6 \pm 0.2$ & $75.3 \pm 0.4$ \\
3-Augment + \schemename & \xmark & \cmark & $81.4 \pm 0.1$ & $81.1 \pm 0.4$ & $79.8 \pm 0.1$ \\
& & & \grntxt{$+2.3$} & \grntxt{$+3.5$} & \grntxt{$+4.5$} \\
\midrule
RandAugment & \cmark & \cmark & $80.1 \pm 0.1$ & $81.9 \pm 0.3$ & $79.3 \pm 2.3$ \\
RandAugment + \schemename & \cmark & \cmark & $80.0 \pm 0.3$ & $81.9 \pm 0.2$ & $82.4 \pm 0.1$ \\
& & & \gtxt{$-0.1$} & \gtxt{$\pm 0.0$} & \grntxt{$+3.1$} \\
\bottomrule
\end{tabular}
}
\end{minipage}
\hfill
\begin{minipage}[t]{.37\textwidth}
\captionof{table}{ImageNet results of models trained on ImageNet with and without \schemename. \schemename improves the performance of most models, with a larger gain for larger models.}
\label{tab:imagenet-results}
\resizebox{\textwidth}{!}{\begin{tabular}{lccc}
\toprule
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{Accuracy [\%]}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& w/o \schemename & w/ \schemename & \\
\midrule
ViT-S & $79.1\pm0.1$ & $81.4\pm0.1$ & \grntxt{$+2.3$} \\
ViT-B & $77.6\pm0.2$ & $81.1\pm0.4$ & \grntxt{$+3.5$} \\
ViT-L & $75.3\pm0.4$ & $79.8\pm0.1$ & \grntxt{$+4.5$} \\
\midrule
DeiT-S & $80.1 \pm 0.1$ & $80.0\pm0.3$ & \gtxt{$-0.1$} \\
DeiT-B & $81.9 \pm 0.3$ & $81.9\pm0.2$ & \gtxt{$\pm0.0$} \\
DeiT-L & $79.3\pm2.3$ & $82.4\pm0.1$ & \grntxt{$+3.1$} \\
\midrule
Swin-Ti & $77.9\pm0.2$ & $79.7\pm0.1$ & \grntxt{$+1.8$} \\
Swin-S & $79.4\pm0.1$ & $80.6\pm0.1$ & \grntxt{$+1.2$} \\
\midrule
ResNet-50 & $78.3\pm0.1$ & $78.8\pm0.1$ & \grntxt{$+0.5$} \\
ResNet-101 & $79.4\pm0.1$ & $80.4\pm0.1$ & \grntxt{$+1.0$} \\
\bottomrule
\end{tabular}}
\end{minipage}
\end{figure}
% \begin{table}[t]
% \caption{ImageNet results of models trained on ImageNet with and without \schemename. \schemename improves the performance of most models, with a larger gain for larger models.}
% \label{tab:imagenet-results}
% \centering
% \begin{subfigure}{.41\textwidth}
% \resizebox{\textwidth}{!}{\begin{tabular}{lccc}
% \toprule
% \multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{ImageNet Accuracy [\%]}} & \multirow{2.5}{*}{Delta} \\
% \cmidrule(lr){2-3}
% & w/o \schemename & w/ \schemename & \\
% \midrule
% ViT-S & $79.1\pm0.1$ & $81.4\pm0.1$ & \grntxt{$+2.3$} \\
% ViT-B & $77.6\pm0.2$ & $81.1\pm0.4$ & \grntxt{$+3.5$} \\
% ViT-L & $75.3\pm0.4$ & $79.8\pm0.1$ & \grntxt{$+4.5$} \\
% \midrule
% Swin-Ti & $77.9\pm0.2$ & $79.7\pm0.1$ & \grntxt{$+1.8$} \\
% Swin-S & $79.4\pm0.1$ & $80.6\pm0.1$ & \grntxt{$+1.2$} \\
% \bottomrule
% \end{tabular}}
% \end{subfigure}
% \hspace{5pt}
% \begin{subfigure}{.448\textwidth}
% \resizebox{\textwidth}{!}{\begin{tabular}{lccc}
% \toprule
% \multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{ImageNet Accuracy [\%]}} & \multirow{2.5}{*}{Delta} \\
% \cmidrule(lr){2-3}
% & w/o \schemename & w/ \schemename & \\
% \midrule
% DeiT-S & $80.1 \pm 0.1$ & $80.0\pm0.3$ & \gtxt{$-0.1$} \\
% DeiT-B & $81.9 \pm 0.3$ & $81.9\pm0.2$ & \gtxt{$\pm0.0$} \\
% DeiT-L & $79.3\pm2.3$ & $82.4\pm0.1$ & \grntxt{$+3.1$} \\
% \midrule
% ResNet-50 & $78.3\pm0.1$ & $78.8\pm0.1$ & \grntxt{$+0.5$} \\
% ResNet-101 & $79.4\pm0.1$ & $80.4\pm0.1$ & \grntxt{$+1.0$} \\
% \bottomrule
% \end{tabular}}
% \end{subfigure}
% \end{table}
\section{Experiments}
\label{sec:experiments}
We conduct a comprehensive suit of experiments to validate the effectiveness of our approach,
comparing ImageNet training with and without \schemename for 10 different models and 5 data augmentation pipelines.
Furthermore, we assess the impact of using \schemename for pretraining on multiple fine-grained downstream datasets.
Finally, we exploit \schemename's control over the image distribution to quantify model behaviors and biases.
We always report the mean and standard deviation of three independent training runs.
% \begin{itemize}
% \item [1.] Training on RecombiNet
% \item ImageNet results (large)
% \item Ablation (TinyImageNet): Foreground position
% \item Ablation (TinyImageNet): Which background (or part of other ablation table?)
% \item Ablation (TinyImageNet+ImageNet For edge blur): Design decisions: Which infill model, pruning threshold, p$\to$t /t$\to$p, foreground rotation range (?), edge blur, original image probability/schedule, Foreground size
% \item With other Data Augmentations
% \item [2.] More evalution metrics
% \item Background accuracy (how to frame/sell? Background bias?) / Background robustness (= foreground with all background)?
% \item Foreground focus
% \item Position bias
% \item Size bias
% \end{itemize}
We conduct a comprehensive suit of experiments to validate the effectiveness of our approach.
We compare training on \name, the ImageNet instantiation of \schemename, to training on ImageNet for 7 different models.
Furthermore, we assess the impact of using \name for pretraining on multiple fine-grained downstream datasets.
Additionally, we use \schemename's control over the image distribution to quantify some model behaviors and biases.
\subsection{Design Choices of \schemename}
\label{sec:ablation}
We start by ablating the design choices of \schemename.
For this, we revert to TinyImageNet \cite{Le2015}, a subset of ImageNet containing 200 categories with 500 images each, and Tiny\name, a version of \schemename derived from TinyImageNet.
\Cref{tab:ablation} presents the results of these ablations.
\begin{table*}[t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccccccccccc}
\toprule
\multirow{2}{*}{Dataset} & Detect. & Infill & FG. & Augmentation & BG. & BG. & edge & original & \multicolumn{2}{c}{TinyImageNet Accuracy} \\
& prompt & Model & size & Order & strategy & pruning & smoothing & image mixing & ViT-Ti [\%] & ViT-S [\%] \\
\cmidrule(r){1-1} \cmidrule(lr){2-9} \cmidrule(l){10-11}
TinyImageNet & & & & & & & & & $66.1\pm0.5$ & $68.3\pm0.7$ \\
Tiny\name & specific & LaMa \cite{Suvorov2021} & mean & crop$\to$paste$\to$color & same & - & - & \gtxt{-} & $64.6\pm0.5$ & $70.0\pm0.6$ \\
\gtxt{Tiny\name} & \gtxt{specific} & \gtxt{LaMa \cite{Suvorov2021}} & range & \gtxt{crop$\to$paste$\to$color} & \gtxt{same} & \gtxt{-} & \gtxt{-} & \gtxt{-} & $65.5\pm0.4$ & $71.2\pm0.5$ \\
\gtxt{Tiny\name} & general & \gtxt{LaMa \cite{Suvorov2021}} & \gtxt{range} & \gtxt{crop$\to$paste$\to$color} & \gtxt{same} & \gtxt{-} & \gtxt{-} & \gtxt{-} & $66.4\pm0.6$ & $72.9\pm0.6$ \\
\gtxt{Tiny\name} & \gtxt{general} & Att. Eraser \cite{Sun2024} & \gtxt{range} & \gtxt{crop$\to$paste$\to$color} & \gtxt{same} & \gtxt{-} & \gtxt{-} & \gtxt{-} & $67.5\pm1.2$ & $72.4\pm0.5$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & paste$\to$crop$\to$color & \gtxt{same} & \gtxt{-} & \gtxt{-} & \gtxt{-} & $67.1\pm1.2$ & $72.9\pm0.5$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & 1.0 & \gtxt{-} & \gtxt{-} & $67.0\pm1.2$ & $73.0\pm0.3$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & 0.8 & \gtxt{-} & \gtxt{-} & $67.2\pm1.2$ & $72.9\pm0.8$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & 0.6 & \gtxt{-} & \gtxt{-} & $67.5\pm1.0$ & $72.8\pm0.7$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & $\sigma_\text{max} = 2.0$ & \gtxt{-} & $67.2\pm0.4$ & $72.9\pm0.5$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & $\sigma_\text{max} = 4.0$ & \gtxt{-} & $65.9\pm0.5$ & $72.4\pm0.6$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & $p=0.2$ & $69.8\pm0.5$ & $75.0\pm0.3$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & $p=0.33$ & $69.5\pm0.4$ & $75.2\pm1.0$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & $p=0.5$ & $70.3\pm1.0$ & $74.2\pm0.2$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & linear & $70.1\pm0.7$ & $74.9\pm0.8$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & reverse lin. & $67.6\pm0.2$ & $73.2\pm0.3$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & cos & $71.3\pm1.0$ & $75.7\pm0.8$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & $\sigma_\text{max} = 4.0$ & \gtxt{cos} & $70.0\pm0.8$ & $75.5\pm0.7$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & orig. & \gtxt{0.8} & \gtxt{$\sigma_\text{max} = 4.0$} & \gtxt{cos} & $67.2\pm0.9$ & $69.9\pm1.0$ \\
\gtxt{Tiny\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & all & \gtxt{0.8} & \gtxt{$\sigma_\text{max} = 4.0$} & \gtxt{cos} & $70.1\pm0.7$ & $77.5\pm0.6$ \\
\midrule
\name & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & \gtxt{-} & \gtxt{cos} & - & $80.5\pm0.1$ \\
\gtxt{\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & \gtxt{same} & \gtxt{0.8} & $\sigma_\text{max} = 4.0$ & \gtxt{cos} & - & $80.7\pm0.1$ \\
\gtxt{\name} & \gtxt{general} & \gtxt{Att. Eraser \cite{Sun2024}} & \gtxt{range} & \gtxt{paste$\to$crop$\to$color} & all & \gtxt{0.8} & \gtxt{$\sigma_\text{max} = 4.0$} & \gtxt{cos} & - & $81.3\pm0.1$ \\
\bottomrule
\end{tabular}}
\caption{Ablation of design decisions of Tiny\name on TinyImageNet and \name on ImageNet.}
\label{tab:ablation}
\end{table*}
\textbf{Prompt.}
% We present the ablation of our main design decisions in \Cref{tab:ablation}.
First, we evaluate the type of prompt used to detect the foreground object.
Here, the \emph{general} prompt, which contains the class and the more general object category, outperforms only having the class name (\emph{specific}).
\textbf{Inpainting.} Attentive Eraser \cite{Sun2024} produces superior results compared to LaMa \cite{Suvorov2021} (see \Cref{sec:infill-model-comparison} for examples).
% When comparing the infill models, the GAN-based LaMa \cite{Suvorov2021} gets outperformed by the Attentive Eraser \cite{Sun2024}.
\textbf{Foreground size}
% We observe that LaMa's often infills unnatural textures compared to Attentive Eraser.
% The size of foreground objects during training has a significant impact on the performance.
% Here, using the greater variability of the \emph{range} strategy increases the performance by $\approx 1\%$ compared to the \emph{mean} strategy.
significantly impacts performance.
Employing a \emph{range} of sizes during recombination, rather than a fixed \emph{mean} size, boosts accuracy by approximately 1 p.p.
This suggests that the added variability is beneficial.
\textbf{Order of data augmentation.}
% (1) Applying the image crop related augmentations \emph{before} pasting the foreground object and the color-based ones \emph{after} pasting or (2) applying all data augmentations after pasting the foreground object.
% While results are ambiguous, we choose the second strategy, as it improves the performance of ViT-S, although not the one of ViT-Ti.
Applying all augmentations after foreground-background recombination (\emph{paste$\to$crop$\to$color}) slightly improves ViT-S's performance compared to applying crop-related augmentations before pasting (\emph{crop$\to$paste$\to$color}).
For ViT-Ti, the results are ambiguous.
\textbf{Background pruning.}
When it comes to the choice of backgrounds to use, we test two pruning thresholds ($t_\text{prune}$) to exclude backgrounds with excessive inpainting.
% and only use backgrounds with an relative size of the infilled region of at most $t_\text{prune}$ (exclusive).
A threshold of $t_\text{prune}=1.0$ means that we use all backgrounds that are not fully infilled.
% We find that the background pruning does not significantly impact the models' performance.
% We choose $t_\text{prune}=0.8$ for the following experiments to exclude backgrounds that are mostly artificial.
Varying $t_\text{prune}$ has minimal impact.
Therefore, we choose $t_\text{prune} = 0.8$ to exclude predominantly artificial backgrounds.
Similarly, applying edge smoothing to foreground masks with Gaussian blurring actually hurts performance on Tiny\name, but slightly improves it on \name.
% One of the most important design decisions is the mixing of the original dataset with \name.
\textbf{Mixing} \name with the original ImageNet data proves crucial.
While constant and linear mixing schedules improve performance over no mixing by $2-3$ p.p. compared to only using Tiny\name, the cosine annealing schedule yields the best results, boosting accuracy by another $0.5-1$ p.p.
\textbf{Background strategy.}
Another point is the allowed choice of background image for each foreground object.
% We evaluate three different strategies.
% (1) Picking the background from which that specific foreground was originally extracted.
% The major difference to ImageNet when using this setup is the variability in size and position of the foreground object.
% (2) Picking a background that originally had a foreground object of the same class in it.
% Here, we have backgrounds where objects of this type can typically appear while also creating a wider variety of samples due to pairing each foreground object with different backgrounds each time.
% (3) Picking any background.
% This choice has the largest variety of backgrounds, but the backgrounds are not semantically related to the foreground object anymore.
% We find in \Cref{fig:bg-strategy} that choosing only a foreground's original background is the worst choice.
We compare using the original background, a background from the same class, and any background.
These strategies go from low diversity and high shared information content between the foreground and background to high diversity and low shared information content.
For \emph{ViT-Ti}, the latter two strategies perform comparably, while \emph{ViT-S} benefits from the added diversity of using any background.
The same is true when training on the full (ImageNet) version of \name.
\begin{figure}
\centering
\includegraphics[width=.7\columnwidth]{img/bates.pdf}
\caption{Plot of the probability distribution function (PDF) of the extended Bates distribution for different parameters $\eta$. Higher values of $\eta$ concentrate the distribution around the center.}
\label{fig:bates-pdf}
\end{figure}
\begin{table}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{ccccccc}
\toprule
\multirow{2.5}{*}{\makecell{Training Set/ \\ Bates Parameter}} & \multirow{2.5}{*}{TIN} & \multicolumn{5}{c}{Tiny\name} \\
\cmidrule(l){3-7}
& & $\eta=-3$ & $-2$ & $1/-1$ & $2$ & $3$ \\
\midrule
TinyImageNet & 68.9 & 60.5 & 60.2 & 60.8 & 62.6 & 63.1 \\
$\eta=-3$ & 71.3 & 79.3 & 79.5 & 79.1 & 79.3 & 79.1 \\
$\eta=-2$ & 71.5 & 80.0 & 78.7 & 79.3 & 79.1 & 78.8 \\
$\eta=1/-1$ & 72.3 & 79.5 & 78.9 & 80.2 & 79.7 & 80.4 \\
$\eta=2$ & 71.3 & 78.2 & 77.8 & 79.1 & 79.6 & 79.9 \\
$\eta=3$ & 71.4 & 77.2 & 76.9 & 78.6 & 79.6 & 79.7 \\
\bottomrule
\end{tabular}}
\caption{Accuracy of ViT-S trained on TinyImageNet (TIN) and Tiny\name with different foreground position distributions by varying the parameter of a Bates distribution $\eta$.
The best performance is achieved using the uniform distribution ($\eta=1$).}
\end{table}
\textbf{Foreground position.}
Finally, we analyze the foreground object's positioning in the image.
We utilize an extended Bates distribution to sample the position of the foreground object.
The Bates distribution~\cite{Bates1955} with parameter $\eta \geq 1$ is the mean of $\eta$ independent uniformly distributed random variables \cite{Jonhson1995}.
Therefore, the larger $\eta$, the more concentrated the distribution is around the center.
We extend this concept to $\eta \leq -1$ by defining ${X \sim \text{Bates}(\eta) :\Leftrightarrow s(X) \sim \text{Bates}(-\eta)}$ for $\eta \leq 1$ with $s$ being the sawtooth function on $[0, 1]$:
\begin{align}
s(x) = \begin{cases}
x + 0.5 & \text{if } 0 < x < 0.5 \\
x - 0.5 & \text{if } 0.5 \leq x \leq 1
\end{cases}
\end{align}
Note that $s \circ s = \id$ on $[0, 1]$.
This way, distributions with $\eta \leq -1$ are more concentrated around the borders.
$\eta = 1$ and $\eta = -1$ both correspond to the uniform distribution.
The PDF of this extended Bates distribution is visualized in \Cref{fig:bates-pdf}.
When sampling more towards the center of the image, the difficulty of the task is reduced, which then reduces the performance on TinyImageNet.
This is reflected in the performance when evaluating on Tiny\name with $\eta=2$ and $\eta=3$ compared to $\eta=-1/1$.
We observe a similar reduction for $\eta < -1$.
This experiment is conducted using the LaMa infill model.
\begin{table}
\centering
\small
\begin{tabular}{lccc}
\toprule
Dataset & Classes & \makecell{Training \\ Images} & \makecell{Validation \\ Images} \\
\midrule
TinyImageNet & 200 & 100,000 & 10,000 \\
Tiny\name & 200 & 99,404 & 9,915 \\
ImageNet & 1,000 & 1,281,167 & 50,000 \\
\name & 1,000 & 1,274,557 & 49,751 \\
\bottomrule
\end{tabular}
\caption{Dataset statistics for TinyImageNet, Tiny\name, ImageNet, and \name. For \name and Tiny\name we report the number of foreground/background pairs.}
\label{tab:dataset-stats}
\end{table}
After fixing the optimal design parameters in \Cref{tab:ablation} (last row), we construct the full \name dataset using the entire ImageNet dataset.
\Cref{tab:dataset-stats} compares the dataset statistics of ImageNet and \name.
% The slightly lower number of images in \name is due to \emph{Grounded SAM} returning no or invalid detections for some images.
The slightly reduced image count in \name is due to instances where Grounded SAM failed to produce valid object detections.
\subsection{Image Classification Results}
\textbf{ImageNet training.}
\Cref{tab:imagenet-pipelines} analyzes the effect of \schemename under different data augmentation pipelines:
A \emph{basic} pipeline with RandomResizedCrop, Flip and ColorJitter, the \emph{3-Augment} pipeline from \cite{Touvron2022,Nauen2025} that also includes Grayscale, Solarization and GaussianBlur, as well as the widely used \emph{RandAugment}~\cite{Cubuk2020} based pipeline from DeiT~\cite{Touvron2021b}.
Additionally, we include MixUp~\cite{Zhang2018a} and CutMix~\cite{Yun2019} augmentations.
% We also include Mixup and CutMix.
We find that the effectiveness of \schemename depends on the interplay between model capacity and baseline augmentation strength.
When the baseline augmentation is weak or moderate, \schemename consistently improves ImageNet accuracy, with gains increasing for larger ViT models (up to $+6.0$ p.p.\ for ViT-B).
As the augmentation pipeline becomes stronger (e.g., RandAugment with MixUp and CutMix), ImageNet improvements diminish for smaller models, indicating that the baseline augmentation already saturates their capacity.
Importantly, even in cases where ImageNet accuracy does not improve, we consistently observe gains during downstream fine-tuning (see \Cref{tab:downstream-results}), suggesting that \schemename enhances representation quality beyond what is reflected by ImageNet accuracy.
\Cref{tab:imagenet-results} additionally compares performance of different model architectures.
ViT~\cite{Dosovitskiy2021}, Swin~\cite{Liu2021} and ResNet~\cite{He2016} (representing CNNs) are trained using the ``3-augment'' strategy, while DeiT~\cite{Touvron2021b} is trained using the ``RandAugment'' strategy.
Notably, \schemename improves performance across all tested architectures, including the ResNet models, % (up to $1$ p.p.),
demonstrating benefits beyond Transformers.
% We find that \schemename's improvements counteract the drop in performance for increasing model sizes.
% Without \schemename this drop is $3.8$ p.p. (ViT-S to L), while with \schemename it is reduced to $1.6$ p.p.
% For DeiT there is a drop of $0.8$ p.p. from small to large while when using \schemename there is a \emph{gain} of $2.4$ p.p.
\begin{table}[t]
\caption{Downstream accuracy in percent when finetuning on other datasets. Models are pretrained on ImageNet with and without \schemename. Pretraining using \schemename increases transformer downstream accuracy.
% on all datasets.
}
\label{tab:downstream-results}
\begin{subfigure}{.48\columnwidth}
\resizebox{\textwidth}{!}{\begin{tabular}{lcccccc}
\begin{table}
\centering
\begin{tabular}{lccc}
\toprule
Model & \schemename & Aircraft & Cars & Flowers & Food & Pets \\
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{ImageNet Accuracy \\ when trained on}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& ImageNet & \name & \\
\midrule
ViT-S & \xmark & $72.4\pm1.0$ & $89.8\pm0.3$ & $94.5\pm0.2$ & $89.1\pm0.1$ & $93.8\pm0.2$ \\
ViT-S & \cmark & $78.6\pm0.5$ & $92.2\pm0.2$ & $95.5\pm0.2$ & $89.6\pm0.1$ & $94.5\pm0.2$ \\
& & \grntxt{$+6.2$} & \grntxt{$+2.4$} & \grntxt{$+1.0$} & \grntxt{$+0.5$} & \grntxt{$+0.7$} \\
ViT-S & $79.1\pm0.1$ & $81.4\pm0.1$ & \grntxt{+2.3} \\
ViT-B & $77.6\pm0.2$ & $81.1\pm0.4$ & \grntxt{+3.5} \\
ViT-L & $75.3\pm0.4$ & $79.8\pm0.1$ & \grntxt{+4.5} \\
\midrule
ViT-B & \xmark & $71.7\pm0.5$ & $90.0\pm0.2$ & $94.8\pm0.4$ & $89.8\pm0.2$ & $94.1\pm0.4$ \\
ViT-B & \cmark & $79.0\pm2.2$ & $93.3\pm0.1$ & $ 96.5\pm0.1$ & $90.9\pm0.1$ & $95.1\pm0.4$ \\
& & \grntxt{$+7.3$} & \grntxt{$+3.3$} & \grntxt{$+1.7$} & \grntxt{$+1.1$} & \grntxt{$+1.0$} \\
Swin-Ti & $77.9\pm0.2$ & $79.7\pm0.1$ & \grntxt{+1.8} \\
Swin-S & $79.4\pm0.1$ & $80.6\pm0.1$ & \grntxt{+1.2} \\
\midrule
ViT-L & \xmark & $72.1\pm1.0$ & $88.8\pm0.3$ & $94.4\pm0.3$ & $90.1\pm0.2$ & $94.2\pm0.4$ \\
ViT-L & \cmark & $77.6\pm1.2$ & $89.1\pm0.2$ & $96.6\pm0.1$ & $91.3\pm0.1$ & $95.1\pm0.1$ \\
& & \grntxt{$+5.5$} & \grntxt{$+0.3$} & \grntxt{$+2.2$} & \grntxt{$+1.2$} & \grntxt{$+0.9$} \\
\midrule
Swin-Ti & \xmark & $77.0\pm0.1$ & $91.3\pm0.6$ & $95.9\pm0.1$ & $90.0\pm0.2$ & $94.2\pm0.1$ \\
Swin-Ti & \cmark & $81.1\pm0.8$ & $92.8\pm0.4$ & $96.2\pm0.1$ & $90.4\pm0.3$ & $94.8\pm0.5$ \\
& & \grntxt{$+4.1$} & \grntxt{$+2.5$} & \grntxt{$+0.3$} & \grntxt{$+0.4$} & \grntxt{$+0.6$} \\
\midrule
Swin-S & \xmark & $75.7\pm1.4$ & $91.0\pm0.3$ & $95.9\pm0.5$ & $91.1\pm0.2$ & $94.4\pm0.1$ \\
Swin-S & \cmark & $81.4\pm0.2$ & $93.1\pm0.2$ & $96.3\pm0.3$ & $91.2\pm0.2$ & $94.9\pm0.3$ \\
& & \grntxt{$+5.7$} & \grntxt{$+2.1$} & \grntxt{$+1.4$} & \gtxt{$+0.1$} & \grntxt{$+0.5$} \\
ResNet-50 & $78.3\pm0.1$ & $78.8\pm0.1$ & \grntxt{+0.5} \\
ResNet-101 & $79.4\pm0.1$ & $80.4\pm0.1$ & \grntxt{+1.0} \\
\bottomrule
\end{tabular}}
\end{subfigure}
\hfill
\begin{subfigure}{.505\columnwidth}
\resizebox{\textwidth}{!}{\begin{tabular}{lcccccc}
\toprule
Model & \schemename & Aircraft & Cars & Flowers & Food & Pets \\
\midrule
DeiT-S & \xmark & $75.3\pm0.4$ & $91.1\pm0.2$ & $94.8\pm0.4$ & $89.2\pm0.2$ & $92.4\pm0.2$ \\
DeiT-S & \cmark & $76.8\pm0.8$ & $91.9\pm0.2$ & $95.2\pm0.3$ & $89.1\pm0.2$ & $92.3\pm0.4$ \\
& & \grntxt{$+1.5$} & \grntxt{$+0.8$} & \grntxt{$+0.4$} & \gtxt{$-0.1$} & \gtxt{$-0.1$} \\
\midrule
DeiT-B & \xmark & $77.0\pm1.2$ & $92.9\pm0.2$ & $96.1\pm0.2$ & $91.2\pm0.1$ & $93.3\pm0.4$ \\
DeiT-B & \cmark & $79.3\pm0.3$ & $93.1\pm0.1$ & $96.4\pm0.2$ & $91.3\pm0.1$ & $93.3\pm0.1$ \\
& & \grntxt{$+2.3$} & \gtxt{$+0.2$} & \grntxt{$+0.3$} & \gtxt{$+0.1$} & \gtxt{$\pm0.0$} \\
\midrule
DeiT-L & \xmark & $72.8\pm5.5$ & $92.8\pm1.0$ & $95.8\pm1.5$ & $90.5\pm2.6$ & $92.4\pm2.0$ \\
DeiT-L & \cmark & $78.8\pm0.8$ & $93.8\pm0.2$ & $97.0\pm0.2$ & $92.0\pm0.2$ & $93.5\pm0.2$ \\
& & \grntxt{$+6.0$} & \grntxt{$+1.0$} & \grntxt{$+1.2$} & \grntxt{$+1.5$} & \grntxt{$+1.1$} \\
\midrule
ResNet-50 & \xmark & $78.2\pm0.5$ & $89.8\pm0.2$ & $91.7\pm0.4$ & $84.4\pm0.2$ & $93.7\pm0.3$ \\
ResNet-50 & \cmark & $80.3\pm0.4$ & $90.4\pm0.2$ & $91.7\pm0.2$ & $84.5\pm0.2$ & $93.7\pm0.3$ \\
& & \grntxt{$+2.1$} & \grntxt{$+0.6$} & \gtxt{$\pm0.0$} & \gtxt{$+0.1$} & \gtxt{$\pm0.0$} \\
\midrule
ResNet-101 & \xmark & $78.4\pm0.6$ & $90.3\pm0.1$ & $91.2\pm0.5$ & $86.0\pm0.2$ & $94.3\pm0.2$ \\
ResNet-101 & \cmark & $81.4\pm0.5$ & $91.3\pm0.1$ & $92.9\pm0.2$ & $86.3\pm0.1$ & $94.0\pm0.3$ \\
& & \grntxt{$+3.0$} & \grntxt{$+1.3$} & \grntxt{$+1.7$} & \grntxt{$+0.3$} & \textcolor{red}{$-0.3$} \\
\bottomrule
\end{tabular}}
\end{subfigure}
\end{tabular}
\caption{ImageNet results of models trained on \name and on ImageNet directly. \name improves the performance of all models in our test.}
\label{tab:imagenet-results}
\end{table}
\textbf{Downstream tasks.} To assess the transferability of \schemename-trained models, we finetune models pretrained on ImageNet with and without \schemename on five fine-grained datasets:
\Cref{tab:imagenet-results} compares the ImageNet performance of models trained on \name and ones trained directly on ImageNet.
We adopt the training setup of \cite{Nauen2023} and \cite{Touvron2022} (details in \Cref{sec:training_setup}) for training ViT \cite{Dosovitskiy2021}, Swin \cite{Liu2021} and ResNet \cite{He2016} models.
Notably, \name improves performance across all tested architectures, including the ResNet models (up to $1$ p.p.), demonstrating benefits beyond Transformers.
For Transformer models, we observe improvements from $1.2$ p.p. to $4.5$ p.p.
This improvement is more substantial for the larger models, with ViT-L gaining $4.5$ p.p. in accuracy.
\name's improvements mostly counteract the drop in performance due to overfitting for large models.
When training on ImageNet, this drop is $3.8$ p.p. from ViT-S to ViT-L, while for \name it is reduced to $1.6$ p.p.
\begin{table}
\centering
\resizebox{\columnwidth}{!}{\begin{tabular}{lccccc}
\toprule
Model & Aircraft & Cars & Flowers & Food & Pets \\
\midrule
ViT-S @ ImageNet & $72.4\pm1.0$ & $89.8\pm0.3$ & $94.5\pm0.2$ & $89.1\pm0.1$ & $93.8\pm0.2$ \\
ViT-S @ \name & $78.6\pm0.5$ & $92.2\pm0.2$ & $95.5\pm0.2$ & $89.6\pm0.1$ & $94.5\pm0.2$ \\
& \grntxt{+6.2} & \grntxt{+2.4} & \grntxt{+1.0} & \grntxt{+0.5} & \grntxt{+0.7} \\
\cmidrule(r){1-1}
ViT-B @ ImageNet & $71.7\pm0.5$ & $90.0\pm0.2$ & $94.8\pm0.4$ & $89.8\pm0.2$ & $94.1\pm0.4$ \\
ViT-B @ \name & $79.0\pm2.2$ & $93.3\pm0.1$ & $ 96.5\pm0.1$ & $90.9\pm0.1$ & $95.1\pm0.4$ \\
& \grntxt{+7.3} & \grntxt{+3.3} & \grntxt{+1.7} & \grntxt{+1.1} & \grntxt{+1.0} \\
\cmidrule(r){1-1}
ViT-L @ ImageNet & $72.1\pm1.0$ & $88.8\pm0.3$ & $94.4\pm0.3$ & $90.1\pm0.2$ & $94.2\pm0.4$ \\
ViT-L @ \name & $77.6\pm1.2$ & $89.1\pm0.2$ & $96.6\pm0.1$ & $91.3\pm0.1$ & $95.1\pm0.1$ \\
& \grntxt{+5.5} & \grntxt{+0.3} & \grntxt{+2.2} & \grntxt{+1.2} & \grntxt{+0.9} \\
\midrule
Swin-Ti @ ImageNet & $77.0\pm0.1$ & $91.3\pm0.6$ & $95.9\pm0.1$ & $90.0\pm0.2$ & $94.2\pm0.1$ \\
Swin-Ti @ \name & $81.1\pm0.8$ & $92.8\pm0.4$ & $96.2\pm0.1$ & $90.4\pm0.3$ & $94.8\pm0.5$ \\
& \grntxt{+4.1} & \grntxt{+2.5} & \grntxt{+0.3} & \grntxt{+0.4} & \grntxt{+0.6} \\
\cmidrule(r){1-1}
Swin-S @ ImageNet & $75.7\pm1.4$ & $91.0\pm0.3$ & $95.9\pm0.5$ & $91.1\pm0.2$ & $94.4\pm0.1$ \\
Swin-S @ \name & $81.4\pm0.2$ & $93.1\pm0.2$ & $96.3\pm0.3$ & $91.2\pm0.2$ & $94.9\pm0.3$ \\
& \grntxt{+5.7} & \grntxt{+2.1} & \grntxt{+1.4} & \grntxt{+0.1} & \grntxt{+0.5} \\
\midrule
ResNet-50 @ ImageNet & $78.2\pm0.5$ & $89.8\pm0.2$ & $91.7\pm0.4$ & $84.4\pm0.2$ & $93.7\pm0.3$ \\
ResNet-50 @ \name & $80.3\pm0.4$ & $90.4\pm0.2$ & $91.7\pm0.2$ & $84.5\pm0.2$ & $93.7\pm0.3$ \\
& \grntxt{+2.1} & \grntxt{+0.6} & \gtxt{$\pm$0} & \grntxt{+0.1} & \gtxt{$\pm$0} \\
\cmidrule(r){1-1}
ResNet-101 @ ImageNet & $78.4\pm0.6$ & $90.3\pm0.1$ & $91.2\pm0.5$ & $86.0\pm0.2$ & $94.3\pm0.2$ \\
ResNet-101 @ \name & $81.4\pm0.5$ & $91.3\pm0.1$ & $92.9\pm0.2$ & $86.3\pm0.1$ & $94.0\pm0.3$ \\
& \grntxt{+3.0} & \grntxt{+1.3} & \grntxt{+1.7} & \grntxt{+0.3} & \textcolor{red}{-0.3} \\
\bottomrule
\end{tabular}}
\caption{Downstream accuracy in percent when finetuning on other datasets. Models were pretrained on \name and ImageNet. Pretraining on \name increases Transformer downstream accuracy on all datasets.}
\end{table}
To assess the transferability of \name-trained models, we finetune models pretrained on ImageNet and \name on five fine-grained datasets:
FGVC-Aircraft \cite{Maji2013}, Stanford Cars~\cite{Dehghan2017}, Oxford Flowers \cite{Nilsback2008}, Food-101 \cite{Kaur2017}, and Oxford-IIIT Pets \cite{Parkhi2012}.
% While for ResNets, the performance of both training datasets is about the same,
In \Cref{tab:downstream-results} we see transformer accuracies improve on all these datasets by up to 7.3 p.p.
% and a reduction of error rate of up to $39.3\%$.
% Notably, training with \name increases the downstream performance of DeiT-S and DeiT-B, even though the ImageNet results were the same.
% This demonstrates that the improved representations from training on \name translate to superior performance beyond gains from better ImageNet performance.
Notably, training with \schemename boosts the downstream performance of DeiT-S and DeiT-B, despite similar ImageNet accuracy.
This shows, that the improved representations from training with \schemename translate to gains beyond better ImageNet scores.
% not only on ImageNet, but also on fine-grained image classification tasks.
While for ResNets, the performance of both training datasets is about the same, for every Transformer, we see the accuracy improve on all downstream dataset by up to 7.3 p.p. and a reduction of error rate of up to $39.3\%$.
In summary, these results demonstrate that the improved representation learning achieved by training on \name translates to superior performance not only on ImageNet, but also on a variety of fine-grained image classification tasks.
\begin{table}[t]
\caption{Evaluation of models trained on ImageNet with and without \schemename. \schemename generally increases models' robustness to different image distribution shifts. Note that ViT-S \emph{with} \schemename outperforms DeiT-S, the only model where \schemename does not increase robustness.}
\label{tab:robustness-datasets}
\begin{subfigure}{.485\textwidth}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccccc}
\subsection{Further Model Evaluation}
% Additional to just using \name for training, its special properties and posibilities for adjustment of the data distribution make it a valuable tool for evaluating other model properties and biases.
Beyond its use for training, \name's unique properties and controlled data generation capabilities make it a powerful tool for analyzing model behavior and biases.
\paragraph*{Background Robustness}
\begin{table}
\centering
\begin{tabular}{lccc}
\toprule
Model & w/ \schemename & IN-Hard & IN-A & IN-C & IN-R & IN-V2 \\
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{Background Robustness \\ when trained on}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& ImageNet & \name & \\
\midrule
ViT-S & \xmark & $18.1 \pm 0.6$ & $18.8 \pm 0.2$ & $44.7 \pm 0.8$ & $41.6 \pm 0.6$ & $67.3 \pm 0.4$ \\
ViT-S & \cmark & $21.0 \pm 0.4$ & $26.5 \pm 0.4$ & $52.6 \pm 0.6$ & $49.8 \pm 0.3$ & $70.6 \pm 0.1$ \\
& & \grntxt{$+2.9$} & \grntxt{$+7.7$} & \grntxt{$+7.9$} & \grntxt{$+8.1$} & \grntxt{$+3.3$} \\
ViT-S & $0.73\pm0.01$ & $0.99\pm0.01$ & \grntxt{+0.26} \\
ViT-B & $0.72\pm0.01$ & $1.00\pm0.01$ & \grntxt{+0.28} \\
ViT-L & $0.70\pm0.01$ & $1.00\pm0.01$ & \grntxt{+0.30} \\
\midrule
ViT-B & \xmark & $17.0 \pm 0.4$ & $15.8 \pm 0.7$ & $40.4 \pm 0.8$ & $38.4 \pm 0.7$ & $65.1 \pm 0.6$ \\
ViT-B & \cmark & $22.0 \pm 0.9$ & $31.9 \pm 1.5$ & $51.6 \pm 1.8$ & $48.7 \pm 1.7$ & $70.3 \pm 0.9$ \\
& & \grntxt{$+5.0$} & \grntxt{$+16.0$} & \grntxt{$+11.2$} & \grntxt{$+10.3$} & \grntxt{$+5.2$} \\
Swin-Ti & $0.72\pm0.01$ & $1.00\pm0.01$ & \grntxt{+0.28} \\
Swin-S & $0.72\pm0.01$ & $1.00\pm0.01$ & \grntxt{+0.28} \\
\midrule
ViT-L & \xmark & $15.6 \pm 0.4$ & $11.3 \pm 0.9$ & $38.4 \pm 1.0$ & $36.8 \pm 0.8$ & $61.6 \pm 0.8$ \\
ViT-L & \cmark & $20.6 \pm 0.1$ & $30.4 \pm 0.5$ & $48.2 \pm 0.7$ & $46.0 \pm 0.4$ & $68.7 \pm 0.3$ \\
& & \grntxt{$+5.0$} & \grntxt{$+19.0$} & \grntxt{$+9.8$} & \grntxt{$+9.3$} & \grntxt{$+7.1$} \\
\midrule
Swin-Ti & \xmark & $16.2 \pm 0.4$ & $15.0 \pm 0.3$ & $36.0 \pm 0.8$ & $36.6 \pm 0.2$ & $65.5 \pm 0.4$ \\
Swin-Ti & \cmark & $18.3 \pm 0.3$ & $20.3 \pm 0.4$ & $41.4 \pm 0.8$ & $41.4 \pm 0.2$ & $68.2 \pm 0.4$ \\
& & \grntxt{$+2.2$} & \grntxt{$+5.4$} & \grntxt{$+5.4$} & \grntxt{$+4.8$} & \grntxt{$+2.7$} \\
\midrule
Swin-S & \xmark & $18.2 \pm 0.3$ & $19.4 \pm 0.3$ & $39.0 \pm 0.7$ & $39.1 \pm 0.2$ & $67.5 \pm 0.1$ \\
Swin-S & \cmark & $20.5 \pm 0.1$ & $27.7 \pm 0.4$ & $45.6 \pm 0.8$ & $44.1 \pm 0.3$ & $69.6 \pm 0.1$ \\
& & \grntxt{$+2.2$} & \grntxt{$+8.4$} & \grntxt{$+6.6$} & \grntxt{$+5.0$} & \grntxt{$+2.2$} \\
ResNet-50 & $0.79\pm0.01$ & $0.99\pm0.01$ & \grntxt{+0.20} \\
ResNet-101 & $0.79\pm0.01$ & $1.00\pm0.01$ & \grntxt{+0.21} \\
\bottomrule
\end{tabular}
}
\end{subfigure}
\hfill
\begin{subfigure}{.505\textwidth}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccccccc}
\toprule
Model & w/ \schemename & IN-Hard & IN-A & IN-C & IN-R & IN-V2 \\
\midrule
DeiT-S & \xmark & $19.5 \pm 0.2$ & $18.4 \pm 0.3$ & $58.8 \pm 0.7$ & $43.0 \pm 0.1$ & $68.8 \pm 0.2$ \\
DeiT-S & \cmark & $18.5 \pm 0.5$ & $17.3 \pm 1.0$ & $57.0 \pm 0.9$ & $43.8 \pm 0.2$ & $68.7 \pm 0.6$ \\
& & \rdtxt{$-1.0$} & \rdtxt{$-1.1$} & \rdtxt{$-1.8$} & \grntxt{$+0.8$} & \gtxt{$-0.1$} \\
\midrule
DeiT-B & \xmark & $22.6 \pm 0.2$ & $26.0 \pm 0.2$ & $62.1 \pm 1.0$ & $45.6 \pm 1.9$ & $70.6 \pm 0.9$ \\
DeiT-B & \cmark & $22.6 \pm 0.2$ & $25.0 \pm 0.3$ & $62.8 \pm 0.6$ & $47.7 \pm 0.8$ & $70.8 \pm 0.5$ \\
& & \gtxt{$\pm 0.0$} & \rdtxt{$-1.0$} & \grntxt{$+0.8$} & \grntxt{$+2.0$} & \gtxt{$+0.2$} \\
\midrule
DeiT-L & \xmark & $21.2 \pm 2.0$ & $20.2 \pm 3.4$ & $59.3 \pm 4.3$ & $41.3 \pm 2.7$ & $66.9 \pm 2.8$ \\
DeiT-L & \cmark & $23.4 \pm 0.3$ & $28.8 \pm 2.0$ & $63.4 \pm 0.7$ & $47.8 \pm 0.6$ & $71.6 \pm 0.5$ \\
& & \grntxt{$+2.2$} & \grntxt{$+8.7$} & \grntxt{$+4.1$} & \grntxt{$+6.5$} & \grntxt{$+4.7$} \\
\midrule
ResNet50 & \xmark & $16.1 \pm 0.2$ & $9.7 \pm 0.1$ & $38.0 \pm 1.0$ & $40.5 \pm 0.6$ & $66.8 \pm 0.4$ \\
ResNet50 & \cmark & $17.2 \pm 0.1$ & $10.8 \pm 0.4$ & $41.0 \pm 0.7$ & $43.7 \pm 0.3$ & $67.5 \pm 0.1$ \\
& & \grntxt{$+1.1$} & \grntxt{$+1.1$} & \grntxt{$+3.0$} & \grntxt{$+3.2$} & \grntxt{$+0.7$} \\
\midrule
ResNet101 & \xmark & $18.2 \pm 0.4$ & $14.3 \pm 0.1$ & $41.7 \pm 0.7$ & $42.3 \pm 0.1$ & $67.7 \pm 0.5$ \\
ResNet101 & \cmark & $19.9 \pm 0.2$ & $17.6 \pm 0.5$ & $46.3 \pm 0.6$ & $46.3 \pm 0.3$ & $69.5 \pm 0.3$ \\
& & \grntxt{$+1.7$} & \grntxt{$+3.2$} & \grntxt{$+4.6$} & \grntxt{$+4.0$} & \grntxt{$+1.8$} \\
\bottomrule
\end{tabular}
}
\end{subfigure}
\end{tabular}
\caption{Evaluation of the background robustness of models trained on \name and on ImageNet directly. Training on \name improves the background robustness of all model to $\approx1.00$, meaning the model is indifferent to the choice of background.}
\label{tab:background-robustness}
\end{table}
\subsection{Bias and Robustness Evaluation}
Beyond its use for training, \schemename's unique properties and controlled data generation capabilities make it a powerful tool for analyzing behavior and biases of black-box models.
We exploit this in two complementary ways.
First, we ask whether \schemename-trained models are more robust on \emph{external} ImageNet robustness benchmarks that are not generated by our pipeline.
Second, we use \schemename's fine-grained control for targeted evaluation of specific dimensions of model bias, such as background reliance and center/size bias.
% Together, these experiments allow us to both \emph{probe} and \emph{improve} robustness along clearly defined axes.
% This combination of standard benchmarks and controlled probes allows us to both quantify robustness improvements and attribute them to changes in particular model behaviors.
\textbf{Robustness on External Distribution Shifts.}
\Cref{tab:robustness-datasets} summarizes accuracy on five widely used ImageNet robustness benchmarks: ImageNet-Hard~\cite{Taesiri2023}, ImageNet-A~\cite{Hendrycks2021}, ImageNet-C~\cite{Hendrycks2019}, ImageNet-R~\cite{Hendrycks2021a}, and ImageNetV2~\cite{Recht2019}.
Across ViTs, Swin Transformers, and ResNets, incorporating \schemename during training generally improves robustness to all considered distribution shifts.
For ViTs, the gains are substantial: for example, ViT-B improves from $15.8\%$ to $31.9\%$ accuracy on ImageNet-A ($+16.0$ p.p.) and from $40.4\%$ to $51.6\%$ on ImageNet-C ($+11.2$ p.p.), with similar improvements for ViT-S and ViT-L.
Swin also benefits consistently, with increases of roughly $2$--$8$ p.p. on most benchmarks, and ResNet sees smaller but steady gains (e.g., up to $+4.6$ points on ImageNet-C).
For DeiT, the picture is more nuanced: DeiT-B and DeiT-L still enjoy robustness improvements, whereas DeiT-S exhibits small decreases on several benchmarks.
Interestingly, however, ViT-S trained with \schemename outperforms the DeiT-S baseline.
This suggests that controlled composition can partially close the robustness gap between lightly and heavily regularized models.
Overall, the consistent improvements on corruption-based, natural and hard examples indicate that the compositional invariances induced by \schemename extend beyond the specific foreground/background manipulations used in its construction.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{img/bg_robustness.pdf}
\caption{Evaluation of background robustness on ImageNet + \schemename, ImageNet9~\cite{Xiao2020} and CounterAnimal~\cite{Wang2024f}.
We plot the in-distribution (top of arrow) and the out-of-distribution (bottom of arrow) accuracy when training with and without \schemename.
We annotate each arrow with its length $\Delta$.
Training with \schemename improves the background robustness of all transformers by mostly boosting the out-of-distribution accuracy.
}
\label{fig:background-robustness}
\end{figure*}
\textbf{Background Robustness.}
% By adjusting the background distribution from using a background from an image of the same class as the foreground to using any background, we can evaluate the robustness of models to shifts in the background distribution.
% We assess background robustness by changing the background distribution, comparing accuracy with backgrounds of the same class as the foreground to using any background.
We assess the robustness of models to shifts in the background distribution from a class-related background to any background.
% We define the background robustness coefficient to be the accuracy of a model on \name when using the same class background divided by the accuracy when using any background:
% Background robustness is defined to be the ratio of accuracy on \name with same-class backgrounds to accuracy with any background:
% \begin{align}
% \text{Background Robustness} = \frac{\text{Acc}(\name_\text{all})}{\text{Acc}(\name_\text{same})}
% \end{align}
% It represents the relative drop in performance under a background distribution shift.
\Cref{fig:background-robustness} presents the background robustness results for three datasets: ImageNet with \schemename (all backgrounds vs. backgrounds of same class), ImageNet9~\cite{Xiao2020} (random backgrounds vs. original backgrounds), and CounterAnimal~\cite{Wang2024f} (counter vs. common background).
The top triangle of each arrow represents the in-distribution backgrounds and the bottom triangle represents the out-of-distribution ones.
We follow ImageNet9 and CounterAnimal and assess the background robustness in terms of the accuracy gap when evaluating a model on images of normal background distribution compared to out-of-distribution backgrounds (length of each arrow; $\Delta$).
% When trained on ImageNet, smaller models generally exhibit greater robustness to changes in the background distribution than larger models and ResNet is more robust than the tested Transformer models.
Crucially, \schemename improves the background robustness of all models and across datasets, reducing the background-gap by boosting the performance on the out-of-background-distribution samples more than the in-distribution ones.
We find a similar trend for the Corner-Cases~\cite{Fatima2025} dataset (see supplementary), highlighting the generalization benefits of \schemename to unusual image compositions.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{img/fg_focus.pdf}
\caption{Evaluation of the foreground focus (\Cref{eq:fg-focus}) using GradCam, GradCam++ and IntegratedGradients (IG) of models trained on ImageNet. Training with \schemename improves the foreground focus of almost all models.}
\label{fig:foreground-focus}
\end{figure*}
\textbf{Foreground Focus.}
Leveraging our inherent knowledge of the foreground masks when using \schemename, as well as common XAI techniques~\cite{Selvaraju2016,Chattopadhay2018,Sundararajan2017}, we can evaluate a model's focus on the foreground object.
% I.e. we measure how much the model's decision depends on the foreground.
We can directly evaluate ImageNet-trained models, but this technique can also be extended to other datasets without relying on manually annotated foreground masks.
To evaluate the foreground focus, we employ Grad-CAM \cite{Selvaraju2016}, Grad-CAM++ \cite{Chattopadhay2018} and IntegratedGradients (IG) \cite{Sundararajan2017} to compute the per-pixel importance of an image for the model's prediction.
The foreground focus is defined to be the ratio of the foreground's relative importance to its relative size in the image:
\begin{align} \label{eq:fg-focus}
\text{FG Focus}(\text{img}) = \frac{\text{Area}(\text{img}) \hspace{3pt} \text{Importance}(\text{fg})}{\text{Area}(\text{fg}) \hspace{3pt} \text{Importance}(\text{img})}
Background robustness is defined to be the ratio of accuracy on \name with same-class backgrounds to accuracy with any background:
\begin{align}
\text{Background Robustness} = \frac{\text{Acc}(\name_\text{all})}{\text{Acc}(\name_\text{same})}
\end{align}
If all pixels uniformly receive the same importance value, the foreground focus is one.
The foreground focus of a model is its average focus over all test images.
\Cref{fig:foreground-focus} presents our findings.
Using \schemename significantly increases the foreground focus of ViT, DeiT and ResNet across all XAI metrics.
% I.e. \schemename-trained models base their decision more on the foreground object compared to the background than models trained without \schemename.
% For Swin, the foreground focus stagnates when measured using GradCam and GradCam++, but almost doubles when using IG.
% We hypothesize that Swin's below-uniform foreground focus reported with GradCam is due to its specific implementation for Swin.
We hypothesize Swin's below-uniform foreground focus with GradCam is due to its specific implementation.
It represents the relative drop in performance under a background distribution shift.
\Cref{tab:background-robustness} presents the background robustness of various models.
When trained on ImageNet, smaller models generally exhibit greater robustness to changes in the background distribution than larger models and ResNet is more robust than the tested Transformer models.
Crucially, training on \name instead of ImageNet improves the background robustness of all models to $\approx1.00$, meaning that these models are agnostic to the choice of background and only classify based on the foreground.
These findings highlight the generalization benefits of \name.
\paragraph*{Foreground Focus}
\begin{table}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcccccc}
\toprule
\multirow{4}{*}{Model} & \multicolumn{6}{c}{Foreground Focus when trained on} \\
\cmidrule(l){2-7}
& IN & FN & IN & FN & IN & FN \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \cmidrule(l){6-7}
& \multicolumn{2}{c}{GradCam} & \multicolumn{2}{c}{GradCam++} & \multicolumn{2}{c}{IG} \\
\midrule
ViT-S & $1.2\pm0.1$ & $2.3\pm0.3$ & $1.2\pm0.1$ & $2.1\pm0.4$ & $1.9\pm0.1$ & $2.7\pm0.1$ \\
ViT-B & $1.2\pm0.1$ & $2.4\pm0.7$ & $1.1\pm0.1$ & $2.1\pm0.1$ & $1.7\pm0.1$ & $2.7\pm0.1$ \\
ViT-L & $1.3\pm0.1$ & $1.6\pm0.1$ & $1.1\pm0.1$ & $1.3\pm0.1$ & $1.3\pm0.1$ & $2.6\pm0.1$ \\
\midrule
Swin-Ti & $0.9\pm0.1$ & $0.7\pm0.1$ & $1.0\pm0.3$ & $0.7\pm0.3$ & $2.5\pm01$ & $4.8\pm0.3$ \\
Swin-S & $0.8\pm0.1$ & $0.7\pm0.1$ & $0.7\pm0.1$ & $0.7\pm0.4$ & $2.4\pm0.1$ & $4.6\pm0.3$ \\
\midrule
ResNet-50 & $2.2\pm0.1$ & $2.7\pm0.1$ & $2.0\pm0.1$ & $2.9\pm0.1$ & $3.2\pm0.1$ & $4.9\pm0.2$ \\
ResNet-101 & $2.3\pm0.1$ & $2.8\pm0.1$ & $2.2\pm0.1$ & $3.0\pm0.1$ & $3.2\pm0.1$ & $4.8\pm0.1$ \\
\bottomrule
\end{tabular}}
\caption{Evaluation of the foreground focus using GradCam, GradCam++ and IntegratedGradients of models trained on \name (FN) and on ImageNet (IN) directly. Training on \name improves the foreground focus of almost all models.}
\label{tab:foreground-focus}
\end{table}
Leveraging our inherent knowledge of the foreground masks when using \name, as well as common XAI techniques~\cite{Selvaraju2016,Chattopadhay2018,Sundararajan2017}, we can evaluate a model's focus on the foreground object.
We can directly evaluate ImageNet trained models, but this technique can also be extended to other datasets without relying on manually annotated foreground-masks.
To evaluate the foreground focus, we employ Grad-CAM \cite{Selvaraju2016}, Grad-CAM++ \cite{Chattopadhay2018} or IntegratedGradients (IG) \cite{Sundararajan2017} to compute the per-pixel importance of an image for the model's prediction.
The foreground focus is defined to be the ratio of the foreground's relative importance to its relative size in the image:
\begin{align}
\text{FG Focus}(\text{img}) = \frac{\text{Area}(\text{img}) \hspace{3pt} \text{Importance}(\text{fg})}{\text{Area}(\text{fg}) \hspace{3pt} \text{Importance}(\text{img})}
\end{align}
The foreground focus of a model is its average foreground focus over all test images.
\Cref{tab:foreground-focus} presents our findings.
Training on \name significantly increasees the foreground focus of ViT and ResNet across all metrics used.
For Swin, the foreground focus stagnates when measured using GradCam and GradCam++, but almost doubles when using IG.
% These differences might be due to the way GradCam is calculated for Swin \todo{cite package website where this is from} and the \todo{common critique of GradCam}.
\begin{table}[t]
\caption{
% Evaluation of the center bias.
Accuracy relative to the center accuracy of multiple instantiations of the models when the foreground objects is in different cells of a $3 \times 3$ grid.
We calculate center bias according to \Cref{eq:center-bias}.
Using \schemename significantly reduces models' center bias.}
\label{tab:center-bias}
\begin{subfigure}{.48\columnwidth}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccc}
\toprule
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{Center Bias [\%] when trained}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& w/o \schemename & w/ \schemename \\
\midrule
ViT-S & \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_ImageNet_v3.pdf} & \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-S_RecombNet_all_v3.pdf} \\
& $25.5\pm0.8$ & $22.0\pm0.3$ & \grntxt{$-3.5$} \\
ViT-B & {\includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_ImageNet_v3.pdf}} & \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-B_RecombNet_all_v3.pdf} \\
& $25.4\pm0.4$ & $19.0\pm0.2$ & \grntxt{$-6.4$} \\
ViT-L & \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_ImageNet_v3.pdf} & \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ViT-L_RecombNet_all_v3.pdf} \\
& $24.3\pm1.1$ & $11.7\pm0.7$ & \grntxt{$-12.6$} \\
\midrule
Swin-Ti & {\includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_ImageNet_v3.pdf}} & {\includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-Ti_RecombNet_all_v3.pdf}} \\
& $25.0\pm0.7$ & $16.5\pm0.2$ & \grntxt{$-8.5$} \\
Swin-S & {\includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_ImageNet_v3.pdf}} & {\includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/Swin-S_RecombNet_all_v3.pdf}} \\
& $23.2\pm0.1$ & $15.6\pm0.2$ & \grntxt{$-7.6$} \\
\bottomrule
\end{tabular} }
\end{subfigure}
\hfill
\begin{subfigure}{.497\columnwidth}
\resizebox{\textwidth}{!}{
\begin{tabular}{lccc}
\toprule
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{Center Bias [\%] when trained}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& w/o \schemename & w/ \schemename \\
\midrule
DeiT-S & {\includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_ImageNet_vNone.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_ImageNet_v3.pdf} } & {\includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_fornet_all_linear_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_fornet_all_linear_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-S_fornet_all_linear_v3.pdf}} \\
& $20.4 \pm 0.2$ & $21.2 \pm 0.1$ & \gtxt{$+0.8$} \\
DeiT-B & {\includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_ImageNet_vNone.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_ImageNet_v3.pdf} } & {\includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_fornet_all_cos_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_fornet_all_cos_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-B_fornet_all_cos_v3.pdf}} \\
& $19.0 \pm 0.7$ & $19.0 \pm 0.2$ & \gtxt{$\pm0.0$} \\
DeiT-L & { \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_ImageNet_v3.pdf} } & { \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_fornet_all_cos_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_fornet_all_cos_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/DeiT-L_fornet_all_cos_v3.pdf} } \\
& $21.2 \pm 0.2$ & $18.0 \pm 0.2$ & \grntxt{$-3.2$} \\
\midrule
ResNet50 & {\includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_ImageNet_v3.pdf}} & {\includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet50_RecombNet_all_v3.pdf}} \\
& $26.3\pm0.3$ & $19.7\pm0.3$ & \grntxt{$-6.6$} \\
ResNet101 & {\includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_ImageNet_v3.pdf}} & {\includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_RecombNet_all_v1.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_RecombNet_all_v2.pdf} \includegraphics[width=.08\columnwidth, valign=c]{img/ResNet101_RecombNet_all_v3.pdf}} \\
& $23.0\pm0.3$ & $19.9\pm0.2$ & \grntxt{$-3.1$} \\
\bottomrule
\end{tabular} }
\end{subfigure}
\centering
\includegraphics[width=.5\columnwidth]{img/colorbar_horizontal.pdf}
\paragraph*{Center Bias}
\begin{table}
\centering
\resizebox{\columnwidth}{!}{
\begin{tabular}{lccc}
\toprule
\multirow{2.5}{*}{Model} & \multicolumn{2}{c}{\makecell{Center Bias when trained on}} & \multirow{2.5}{*}{Delta} \\
\cmidrule(lr){2-3}
& ImageNet & \name \\
\midrule
ViT-S & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-S_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-S_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-S_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-S_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-S_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-S_RecombNetAll_v3.pdf}} \\
& $0.255\pm0.008$ & $0.220\pm003$ & \grntxt{-0.035} \\
ViT-B & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-B_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-B_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-B_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-B_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-B_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-B_RecombNetAll_v3.pdf}} \\
& $0.254\pm0.004$ & $0.190\pm0.002$ & \grntxt{-0.064} \\
ViT-L & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-L_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-L_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-L_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ViT-L_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-L_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ViT-L_RecombNetAll_v3.pdf}} \\
& $0.243\pm0.011$ & $0.117\pm0.007$ & \grntxt{-0.126} \\
\midrule
Swin-Ti & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/Swin-Ti_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-Ti_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-Ti_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/Swin-Ti_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-Ti_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-Ti_RecombNetAll_v3.pdf}} \\
& $0.250\pm0.007$ & $0.165\pm0.002$ & \grntxt{-0.085} \\
Swin-S & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/Swin-S_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-S_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-S_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/Swin-S_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-S_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/Swin-S_RecombNetAll_v3.pdf}} \\
& $0.232\pm0.001$ & $0.156\pm002$ & \grntxt{-0.076} \\
\midrule
ResNet50 & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ResNet50_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet50_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet50_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ResNet50_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet50_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet50_RecombNetAll_v3.pdf}} \\
& $0.263\pm0.003$ & $0.197\pm0.003$ & \grntxt{-0.066} \\
ResNet101 & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ResNet101_ImageNet_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet101_ImageNet_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet101_ImageNet_v3.pdf}} & \raisebox{-6pt}{\includegraphics[width=.08\columnwidth]{img/ResNet101_RecombNetAll_v1.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet101_RecombNetAll_v2.pdf} \includegraphics[width=.08\columnwidth]{img/ResNet101_RecombNetAll_v3.pdf}} \\
& $0.230\pm0.003$ & $0.199\pm002$ & \grntxt{-0.031} \\
\bottomrule
\end{tabular} }
\includegraphics[width=.75\columnwidth]{img/colorbar_horizontal.pdf}
\caption{Evaluation of the position bias. We plot the accuracy relative to the center accuracy of multiple instantiations of the models when the foreground objects is in different cells a $3 \times 3$ grid.
Training on \name significantly reduces a models center bias.}
\label{tab:center-bias}
\end{table}
\textbf{Center Bias.}
With \schemename we have unique control over the position of the foreground object in the image.
This lets us quantify the center bias of models trained with and without \schemename.
We divide the image into a $3 \times 3$ grid and evaluate model accuracy when the (scaled-down) foreground object is in each of the $9$ grid cells.
With \name we have unique control over the position of the foreground object in the image.
This lets us quantify the center bias of ImageNet- and \name-trained models.
We divide the image into a $3 \times 3$ grid and evaluate model accuracy when the foreground object is in each of the $9$ grid cells.
Each cell's accuracy is divided by the accuracy in the center cell for normalization, which gives us the relative performance drop when the foreground is in each part of the image.
The center bias is calculated as one minus the average of the minimum performance of a corner cell and the minimum performance of a side cell:
% \begin{align}
% \begin{split}
% & \text{Center Bias} = \\
% & \hspace{7pt} 1 - \frac{\min\limits_{a, b \in \{0, 2\}} \text{Acc}(\text{cell}_{(a, b)}) + \min\limits_{\substack{a=1 \text{ or } b=1 \\ a \neq b}} \text{Acc}(\text{cell}_{(a, b)})}{2 \text{Acc}(\text{cell}_{(1, 1)})}
% \end{split}
% \end{align}
\begin{align} \label{eq:center-bias}
\text{Center Bias} = 1 - \frac{\min\limits_{c \in \text{sides}} \text{Acc}(c) + \min\limits_{c \in \text{corners}} \text{Acc}(c)}{2 \text{Acc}(c_\text{center})}
\begin{align}
\begin{split}
& \text{Center Bias} = \\
& \hspace{7pt} 1 - \frac{\min\limits_{a, b \in \{0, 2\}} \text{Acc}(\text{cell}_{(a, b)}) + \min\limits_{\substack{a=1 \text{ or } b=1 \\ a \neq b}} \text{Acc}(\text{cell}_{(a, b)})}{2 \text{Acc}(\text{cell}_{(1, 1)})}
\end{split}
\end{align}
\Cref{tab:center-bias} visualizes the center bias of three instantiations of each model.
Performance is generally highest in the center and lowest in the four corners.
Performance is generally highest in the center and the center top and bottom and center left and right cells, and lowest in the four corners.
Interestingly, ImageNet-trained models perform slightly better when the foreground object is on the right side of the image, compared to the left side, despite our use of random flipping with a probability of $0.5$ during training.
% Training on \name reduces the center bias of all models by at least half.
Using \schemename significantly reduces center bias across models, with a more uniform performance especially across the middle row.
% On corner-cases (see supplementary) we find that
% Their accuracy is higher in the center left and right cells than in the center top and bottom ones, which is not the case for ImageNet-trained models.
% This demonstrates that \schemename promotes a more uniform spatial attention distribution, counteracting the center-bias of ImageNet.
Thus, \schemename makes the model recognize objects across a wider spatial distribution, counteracting the center-bias of ImageNet.
Training on \name significantly reduces center bias across all models.
This demonstrates that \name promotes a more uniform spatial attention distribution.
Their accuracy is higher in the center left and right cells than in the center top and bottom ones, which is not the case for ImageNet-trained models.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{img/size_bias_wide.pdf}
\caption{Evaluation of the size bias of models trained on ImageNet. We plot the accuracy relative to the accuracy when using the default size ($f_\text{size} = 1.0$).}
\label{fig:size-bias}
\paragraph*{Size Bias}
\begin{figure}
\centering
\includegraphics[width=.9\columnwidth]{img/size_bias.pdf}
\caption{Evaluation of the size bias of models trained on \name. We plot the accuracy relative to the accuracy when using the mean foreground size.}
\label{fig:size-bias}
\end{figure}
\textbf{Size Bias.}
Finally, we evaluate the impact of different sized foreground objects on the accuracy.
Finally, we evaluate the impact of different-sized foreground objects on the accuracy.
For this evaluation, we use the \emph{mean} foreground size strategy.
We introduce a size factor $f_\text{size}$ by which we additionally scale the foreground object before pasting it onto the background.
Results are normalized by the accuracy when using $f_\text{size} = 1.0$.
\Cref{fig:size-bias} shows the size bias curves of models trained with and without \schemename.
Results are again normalized by the accuracy when using the mean foreground size ($f_\text{size} = 1.0$).
\Cref{fig:size-bias} shows the size bias curves of ViT-S and ViT-B when trained on ImageNet and \name.
% When training on \name, the resulting model keeps it's good performance on smaller foreground objects, while models trained on ImageNet fall of faster and lower.
Models trained using \schemename perform better, especially with smaller foreground objects.
%, when ImageNet-trained models exhibit a more rapid performance decline.
Therefore, \schemename-training improves robustness to variations in object scale, especially for larger models.
\subsection{Design Choices of \schemename}
We next analyze key components of \schemename, focusing on three questions: how it compares to simple copy-paste, how background choice affects performance, and how reliably labels are preserved after recomposition.
Additional ablations over variants and hyperparameters are provided in the supplementary material.
\begin{table}[t]
\caption{Comparison of \schemename and simple Copy-Paste methods. We train ViT-S on ImageNet using the same 3-augment data augmentation on top of the copy-paste augmentation.}
\label{tab:copy-paste-comparison}
\centering
\resizebox{.66\columnwidth}{!}{
\begin{tabular}{lcc S[table-format=+2.1,retain-explicit-plus,detect-inline-weight=math,detect-weight=true]}
\toprule
Augmentation & labels & \makecell{ Accuracy [\%]} & {\makecell{Delta \\to Prev.}} \\
\midrule
% Baseline & & $79.1 \pm 0.1$ \\
3-Augment + \textbf{Simple Copy-Paste} & bg & $31.3 \pm 0.6$ & \\
+ mixed labels & fg + bg & $32.0 \pm 0.8$ & +0.7 \\
+ fg labels & fg & $31.6 \pm 0.9$ & -0.4 \\
+ \emph{range} foreground size variation & \gtxt{fg} & $43.0 \pm 1.2$ & \bfseries +11.4 \\
+ infilled backgrounds & \gtxt{fg} & $68.7 \pm 0.2$ & \bfseries +25.7 \\
+ \emph{cos} mixing strategy & \gtxt{fg} & $81.2 \pm 0.1$ & \bfseries +12.5 \\
+ edge smoothing & \gtxt{fg} & $81.3 \pm 0.1$ & +0.1 \\
+ background pruning$=$ \textbf{\schemename} & \gtxt{fg} & $81.4 \pm 0.1$ & +0.1 \\
\bottomrule
\end{tabular}}
\end{table}
\textbf{Comparison to Simple Copy-Paste.}
We compare \schemename to a simple adaption of the Copy-Paste augmentation inspired by \cite{Ge2023,Ghiasi2021,Shermaine2025} in \Cref{tab:copy-paste-comparison}.
Contrary to semantic segmentation we do not have foreground masks available.
Thus, we paste the extracted objects from \textbf{\schemename's segmentation stage} onto normal ImageNet images.
% Since such images do not have straight forward classification labels, we test multiple possibilities.
We observe 3 large jumps in accuracy: (\textbf{1}) From our \emph{range} foreground size variation (+11.4\%), (\textbf{2}) from using our infilled backgrounds instead of images from the dataset (+25.7\%), and (\textbf{3}) from our \emph{cos} mixing strategy with non-augmented images (+12.5\%).
\schemename's changes to the naive copy-paste augmentation are thus imperative for good classification performance.
\begin{figure}[t]
\begin{minipage}[c]{.49\textwidth}
\centering
\includegraphics[width=\textwidth]{img/strategy.pdf}
\captionof{figure}{We compare Original, Same-class, and All-classes background selection using ViT-Ti and ViT-S backbones on TinyImageNet.
Increasing background diversity consistently improves classification accuracy.
}
\label{fig:background-strategy}
\end{minipage}
\hfill
\begin{minipage}[c]{.49\textwidth}
\centering
\includegraphics[width=\textwidth]{img/mask_expansion.pdf}
\captionof{figure}{
We vary the foreground mask area for TinyImageNet by shrinking or expanding masks relative to the original outline and report accuracy when training on $100\%$ augmented samples.
Performance is stable for expanded masks and degrades rapidly after shrinking masks.
}
\label{fig:mask-expansion}
\end{minipage}
\end{figure}
\textbf{Background Choice Strategy.}
\Cref{fig:background-strategy} shows the effect of background selection on TinyImageNet accuracy, where we trade off diversity against context plausibility.
% Using the original inpainted background yields the lowest accuracy, indicating limited regularization from contextual cues.
% Sampling backgrounds from the same class provides a modest but consistent improvement, suggesting that mild context variation encourages robustness while preserving semantic plausibility.
The best performance is achieved by sampling backgrounds from all classes, which introduces substantial context shifts, but leads to the strongest accuracy gains for both ViT-Ti and ViT-S.
Thus, aggressive background diversification is more important than context plausibility and acts as an effective form of context-based regularization rather than introducing harmful noise.
\textbf{Label Integrity.}
% We assess the label integrity of \schemename, i.e., whether object labels remain correct after recombination, by verifying that the intended object is accurately extracted.
% To this end, we leverage the object bounding box annotations provided in the ImageNet validation set.
% Specifically, we compute the \emph{box precision}, defined as the fraction of the predicted mask area that lies within the ground-truth bounding box, obtaining a mean value of $91\%$.
% In addition, we measure the \emph{box-to-box IoU}, computed as the IoU between the tight bounding box enclosing the predicted mask and the tight bounding box of the ground-truth annotation, which yields a high $76.1\%$.
% Qualitative examples of the predicted masks and bounding boxes are provided in the supplementary material.
% We additionally test label integrity under systematic mask perturbations by expanding or shrinking the foreground masks before composition.
% Concretely, starting from the original outline, we erode or dilate the mask such that the foreground area changes by some percentage.
% \Cref{fig:mask-expansion} shows that accuracy is relatively stable for expanded masks, but drops off significantly for eroded masks, consistent with cropping away semantically important object parts.
% This experiment suggests, that \schemename is relatively robust to artifacts from including an object's original background in the foreground mask.
% Overall, these results indicate that the segmentation stage of \schemename reliably isolates the target class object, thereby preserving label correctness after recombination.
To quantify whether recombined images still depict the intended class, we evaluate the segmentation stage of \schemename on ImageNet validation boxes.
Our predicted masks achieve a mean box precision of $91.0\%$ (fraction of mask area inside the ground-truth bounding boxes of the ImageNet validation set) and a high box-to-box IoU of $76.1\%$, indicating that they tightly capture the target object.
Qualitative examples of the predicted masks and bounding boxes are provided in the supplementary material.
We further probe robustness to mask imprecision by eroding or dilating masks such that the foreground area changes by a fixed percentage before composition.
As shown in \Cref{fig:mask-expansion}, accuracy remains stable for expansions but drops sharply under erosion, consistent with removing semantically important object parts.
Together, these results suggest that (\textit{i}) \schemename reliably isolates the target objects and preserves label integrity and that (\textit{ii}) \schemename is robust to artifacts from an object's original background and degrades mainly when the foreground no longer contains the full object.
Models trained on \name maintain better performance even with smaller foreground objects, when ImageNet-trained models exhibit a more rapid performance decline.
Therefore, \name-training improves robustness to variations in object scale.