Reproduce “Towards Purified Multi-Label Test-Time Adaptation of Vision-Language Models” 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
Yiwen Liang, Hui Chen, Yizhe Xiong, Mengyao Lyu, Yuhan Cao, Zijia Lin, Shuaicheng Niu, Sicheng Zhao and 2 more
The research by Liang et al. from Tsinghua University developed PuRF, a purification-driven, cache-based method for multi-label test-time adaptation (MLTTA) of vision-language models. This approach leverages purified region-level visual cues and adaptive cache management to address dominant-label bias and noisy representations. PuRF consistently improved mAP across diverse backbones and datasets, achieving an average mAP gain of 8.56% over ML-TTA and 4.05% over ReTA on ViT-B/32, while maintaining practical efficiency.
Key insights
- Purifying both region-level visual cues and cache representations is crucial for overcoming dominant-label bias and noise in multi-label test-time adaptation.
- Adaptive selection of fine-grained regional evidence, rather than relying solely on global features or all regions, significantly enhances discriminability for multi-label recognition.
- Dynamically managing the cache with episodic and temporal purification strategies ensures class-specific, high-quality prototypes and prevents cache saturation, improving long-term adaptability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.