Reproduce “Microsoft COCO: Common Objects in Context” 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
Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan and 2 more
COCO is a dataset for training and evaluating object recognition in everyday scenes, with labels outlining individual instances. The authors collected contextual photographs and used crowdsourcing to identify categories, locate instances, and create masks. In a preliminary evaluation on 55,000 images across 20 PASCAL VOC categories, COCO-trained DPM averaged 19.1 AP versus 16.9; this does not establish gains across all 91 categories or models.
Key insights
- Crowdsourcing with a multi-stage, quality-controlled pipeline can effectively generate large-scale, high-quality instance segmentation annotations, surpassing the recall of individual expert annotators.
- Emphasizing images with objects in 'non-iconic' views and dense scenes (average 7.7 instances/image) forces models to learn more generalized and context-aware object representations.
- Precise 2D localization through instance segmentation masks provides a more accurate and challenging benchmark for evaluating object recognition algorithms compared to bounding boxes alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.