Researchers have introduced a new benchmark for webly supervised multi-label recognition, a field that leverages freely available web images to reduce reliance on costly manual annotations. This benchmark, named WS-MLR, includes two datasets, Web-COCO and Web-Pascal, which contain approximately 300,000 images across 80 and 20 categories respectively. To further advance the field, a Dual-Branch Multi-Label Contrastive Learning (DBMLCL) framework was proposed, which excels at identifying and correcting noisy labels by learning category-specific representations and their similarities. Experiments show that DBMLCL outperforms existing baseline methods on this new benchmark. AI
IMPACT This work aims to improve the efficiency of training multi-label recognition models by leveraging readily available web data, potentially reducing annotation costs.
RANK_REASON The item is a research paper detailing a new benchmark and a novel framework for a specific machine learning task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
