Reproduce “OutfitTransformer: Learning Outfit Representations for Fashion Recommendation” 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
Rohan Sarkar, Navaneeth Bodla, Mariya I. Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, Gerard Medioni
Researchers from Purdue University and Amazon developed OutfitTransformer, a Transformer-based framework designed to learn unified outfit representations for fashion compatibility prediction (CP) and complementary item retrieval (CIR). This framework achieved improved compatibility prediction AUC scores up to 0.93 and demonstrated robust performance in fill-in-the-blank tasks, reaching 67.10% accuracy.
Key insights
- End-to-end training of the image and text encoders significantly improves compatibility prediction performance, boosting CP-AUC by 9% by enabling the model to learn fashion-specific features.
- Pre-training the model on the compatibility prediction task before fine-tuning for retrieval substantially enhances fill-in-the-blank accuracy, providing a foundational understanding of item relationships.
- The proposed set-wise outfit ranking loss, particularly its hard negative component, and curriculum learning for negative sampling contribute to more stable training and improved discriminative power for complementary item retrieval.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.