Image classifiers are better at cheating than seeing.
Researchers have released ForAug, an open-source augmentation tool that breaks a long-standing shortcut in image classification training. Most large-scale datasets share a quiet problem: objects are usually centered, appear at predictable sizes, and sit against class-specific backgrounds. A model trained on such data learns to lean on those cues instead of the object itself — and falls apart the moment the cues change. ForAug addresses this by splitting each training image into its foreground object and background, then recombining them with different backgrounds, positions, and scales using off-the-shelf segmentation and inpainting models. The method is tested across 10 architectures.
The numbers are hard to dismiss. ForAug lifts ImageNet top-1 accuracy by up to 6 percentage points and pushes fine-grained downstream dataset gains as high as 7.3 points. On standard distribution-shift benchmarks — the tests designed to expose exactly this kind of fragility — accuracy climbs by up to 19 points. That last figure matters because distribution-shift benchmarks are where most published augmentation methods quietly underperform their headlines.
The technique also doubles as a diagnostic: the same foreground-background swap controls let researchers measure how much a model relies on background context, center position, or object size — biases that are easy to overlook and hard to quantify with standard tools. The code and dataset are public on GitHub, which at least removes the usual reproducibility excuse.