Reproduce “VICRegL: Self-Supervised Learning of Local Visual Features” 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
Adrien Bardes, Jean Ponce, Yann LeCun
VICRegL trains global and local image representations together by applying VICReg to pooled features and selected local feature pairs matched by image location or embedding similarity, without segmentation masks or negative pairs. With a frozen ResNet-50, VICRegL at α=0.75 reaches 55.9 mIoU on Pascal VOC versus 47.8 for VICReg, while ImageNet linear classification accuracy is 70.4% versus 71.5%. The frozen-backbone evaluation uses labeled downstream heads; gains show segmentation information is more readily extracted, not that features encode only their mapped objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.