Masked image modeling (MIM) is a promising option for training Vision Transformers among various self-supervised learning (SSL) methods. The essence of MIM lies in token-wise masked token predictions, with targets tokenized from images or generated by pre-trained models such as vision-language models. While tokenizers or pre-trained models are plausible MIM targets, they often offer spatially inconsistent targets even for neighboring tokens, complicating models to learn unified discriminative representations. Our pilot study confirms that addressing spatial inconsistencies has the potential to enhance representation quality. Motivated by the findings, we introduce a novel self-supervision signal called Dynamic Token Morphing (DTM), which dynamically aggregates contextually related tokens to yield contextualized targets. DTM is compatible with various SSL frameworks; we showcase an improved MIM by employing DTM, barely introducing extra training costs. Our experiments on ImageNet-1K and ADE20K demonstrate the superiority of our methods compared with state-of-the-art, complex MIM methods. Furthermore, the comparative evaluation of the iNaturalists and fine-grained visual classification datasets further validates the transferability of our method on various downstream tasks. Code is available at https://github.com/naver-ai/dtm
翻译:暂无翻译