Reproduce “Fashion Outfit Complementary Item Retrieval” 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
Yen-Liang Lin, Son Tran, Larry S. Davis
Researchers from Amazon introduced the Category-based Subspace Attention Network (CSA-Net), a framework for scalable fashion outfit complementary item retrieval. It leverages category-based subspace attention and an outfit ranking loss to achieve 63.73% FITB accuracy and 0.91 AUC for compatibility prediction, and improved Recall@top k metrics by 2-5% over baselines in retrieval tasks on the Polyvore Outfits dataset.
Key insights
- Conditioning subspace attention on item categories (instead of specific item images) is crucial for enabling scalable pre-computation and indexing of embeddings for efficient retrieval.
- A holistic Outfit Ranking Loss, considering the entire outfit context, significantly improves compatibility prediction and retrieval performance over traditional pairwise approaches.
- The model successfully learns to retrieve visually diverse but stylistically compatible items, demonstrating its ability to capture global outfit style rather than just visual similarity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.