SnapMix: Semantically Proportional Mixing for Augmenting Fine-grained Data

Data mixing augmentation has proved effective in training deep models. Recent methods mix labels mainly according to the mixture proportion of image pixels. Due major discriminative information a fine-grained usually resides subtle regions, these tend introduce heavy label noise recognition. We propose Semantically Proportional Mixing (SnapMix) that exploits class activation map (CAM) lessen augmenting data. SnapMix generates target for mixed by estimating its intrinsic semantic composition. This strategy can adapt asymmetric operations and ensure correspondence between synthetic images labels. Experiments show our method consistently outperforms existing mixed-based approaches regardless different datasets or network depths. Further, incorporating mid-level features, proposed achieves top-level performance, demonstrating potential serve as strong baseline