Reproduce “PACO: Parts and Attributes of Common Objects” 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
Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez and 6 more
Introducing PACO, a large-scale dataset, facilitates fine-grained visual understanding by providing joint annotations for object parts and descriptive attributes across 75 common object categories. Benchmarks reveal that current models achieve 17.3 AP for part segmentation and 18.8 AP for object-attribute prediction, significantly outperforming prior open-vocabulary detectors on zero-shot instance retrieval.
Key insights
- The PACO dataset provides an unprecedented resource for developing models capable of compositional scene and object interpretation, connecting vision with descriptive language.
- Part segmentation and attribute prediction, especially at the part-level, represent challenging tasks where even advanced models show substantial room for improvement.
- Zero-shot instance detection using descriptive queries is a significant challenge for current open-vocabulary models, highlighting the need for more robust vision-language integration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.