Title: OA-CutMix: Rectifying the Label Bias in CutMix

Abstract

Although CutMix has emerged as the standard for mixing augmentations, its label assignment mechanism is built on a faulty premise: the assumption that the geometric area of a pasted patch accurately represents its semantic contribution to the composite image. In reality, patches often overlap with background areas, incorrectly attributing label credit to classes whose objects remain invisible. Data indicates that the mean discrepancy between the CutMix label and the actual semantic object area stands at $21.5\%$. Furthermore, in $17\%$ of cases, an image contributes no visible object pixels to the mix yet still receives a non-zero label weight. To address this, we introduce Object-Aware CutMix (OA-CutMix). This method eliminates the bias by substituting the area-based weight with values derived from precomputed segmentation masks, thereby assigning labels based on the proportion of visible object area each image contributes. The core image mixing process remains untouched. Our evaluation of OA-CutMix spans four architectures and six datasets, comparing it against more than ten static and dynamic mixing techniques. OA-CutMix consistently delivers the highest accuracy across all tasks, surpassing even dynamic methods while incurring only a fraction of their training-time costs. The performance gains are most pronounced for small objects, where CutMix’s label bias is most severe. Consequently, simply correcting the label bias is sufficient to match or outperform methods that alter the image mixing algorithm itself.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC