CLIPCleaner: Cleaning Noisy Labels with CLIP

Abstract

Figure 1: The normalized losses distribution of WebVision dataset after one epoch warm-up training, i.e., training with whole dataset and cross-entropy loss. In (a), 'clean'/'noisy' denotes samples been identified as clean/noisy by CLIPCleaner while the 'gray vertical line' denotes the sample selection boundary induced by 'small-loss' mechanism. We show some example images on part 1 in (b) and part 4 in (c) which represents samples identified as 'clean' by 'small-loss' while rejected by CLIPCleaner and vice versa. For example, in (b) we can find that many images with small losses due to its similar color or textures to 'tench' class, thus been wrongly identified as 'clean' by 'small-loss' but been correctly rejected by CLIPCleaner.