Reproduce “Denoising Vision Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiawei Yang, Katie Z Luo, Jiefeng Li, Congyue Deng, Leonidas Guibas, Dilip Krishnan, Kilian Q Weinberger, Yonglong Tian and 1 more
A two-stage Denoising Vision Transformers (DVT) pipeline effectively removes pervasive grid-like artifacts from Vision Transformer feature maps, improving performance on dense prediction tasks such as semantic segmentation (e.g., DINOv2-base mIoU improved to 84.84 on VOC2012) and enhancing unsupervised object discovery.
Key insights
- ViT feature maps can be decomposed into a noise-free semantic component, an input-independent artifact component, and a residual, with positional embeddings identified as the primary source of the artifacts.
- The observed grid-like artifacts are pervasive across various ViT architectures and training paradigms (self-supervised, supervised, multi-modal), and across different model sizes.
- Removing these feature artifacts leads to emergent capabilities, such as a significant improvement in unsupervised object discriminability and discovery.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.