Reproduce “Region-Based Representations Revisited” 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

Michal Shlapentokh-Rothman, Ansel Blume, Yao Xiao, Yuqun Wu, Sethuraman T V, Heyi Tao, Jae Yong Lee, Wilfredo Torres and 2 more

The authors introduce region-based representations that pool pretrained image features within masks from a class-agnostic segmenter, creating vectors for downstream recognition. In matched frozen DINOv2 ViT-L/14 linear-probe tests, regions score 50.2 mIoU on ADE20K versus 43.4 for patches; COCO one-shot retrieval reaches 0.45 mAP versus 0.27 for CLIP-Crop. Results are competitive baselines, and SAM is slow.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.