Reproduce “Knowing Where to Focus: Attention-Guided Alignment for Text-based Person Search” 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
Lei Tan, Weihao Li, Pingyang Dai, Jie Chen, Liujuan Cao, Rongrong Ji
A new Attention-Guided Alignment (AGA) framework enhances text-based person search by dynamically masking semantically important words and enriching low-quality text descriptions. This approach leads to state-of-the-art retrieval performance, achieving a Rank-1 accuracy of 78.36% on CUHK-PEDES, a 1.85% improvement over previous methods.
Key insights
- Dynamically focusing Masked Language Modeling (MLM) on semantically meaningful words, guided by attention mechanisms, significantly improves the efficiency and effectiveness of cross-modal alignment.
- Self-supervised textual data augmentation, through logit-based synonym replacement, can effectively mitigate the negative impact of low-quality and monotonous human-annotated descriptions on VLP model training.
- Combining targeted masking and text enrichment provides a synergistic effect, leading to more robust and accurate representations for text-based person search.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.