Reproduce “Mask-supervised Object-centric Representation Learning with LeJEPA” 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
Jakob Geusen, Ender Konukoglu
The authors introduce object-centric pretraining with separate semantic and instance spaces: it aligns the same object across views while distinguishing instances from one another and context. Copy-paste changes context. With COCO ground-truth masks, a ViT-Base trained on 1,024 images scores 0.646 on DAVIS J&F, versus 0.613 for image-level LeJEPA trained on 118,000 images; this frozen-feature comparison uses unequal training steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.