Reproduce “Object-centric Binding in Contrastive Language-Image Pretraining” 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
Rim Assouel, Pietro Astolfi, Florian Bordes, Michal Drozdzal, Adriana Romero-Soriano
OC-CLIP represents captions as scene graphs and binds their entities to separate visual slots before scoring object matches and relations. Its structured comparison improves attribute and spatial image–caption matching over in-domain-finetuned OpenCLIP: on synthetic seen animal pairs, accuracy is 97% versus 81%; on COCO-spatial, its B-14 model scores 89.7% versus 45.6%. These results concern tested benchmarks and training vocabulary.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.