Reproduce “MulVec: Fine-Grained Role-Aware Matching for Training-Free Zero-Shot Composed Image 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
Zihao Zhang, Dayan Wu, Xinze Liu, Hengjie Zhu, Yiliang Zhu, Ding Wang, Peng Fu, Zheng Lin and 1 more
TAPR introduces an Atomic Edit Factorization framework for training-free composed image retrieval, which explicitly disentangles edit requests into "Target," "Add," "Preserve," and "Remove" factors. This approach achieved state-of-the-art results, improving performance on CIRCO mAP@5 by up to 23.0% over baselines and consistently outperforming on CIRR and FashionIQ datasets.
Key insights
- Explicitly factorizing complex edit requests into atomic components (Target, Add, Preserve, Remove) prevents semantic entanglement and enables more precise verification of each distinct requirement.
- Employing fine-grained multi-vector matching with local visual evidence allows the system to accurately address specific details of an edit that might be overlooked by global similarity scores.
- Assigning distinct, factor-specific operators (e.g., rewarding "Add" evidence, penalizing "Remove" evidence) is crucial for aligning the matching logic with the semantic intent of each edit component.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.