Reproduce “End-to-End Human-Gaze-Target Detection with Transformers” 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
Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, Wei Shen
HGTTR, developed by researchers from Shanghai Jiao Tong University and Baidu Research, introduces an end-to-end Transformer-based framework for human-gaze-target detection. This approach simultaneously predicts all human head locations and their corresponding gaze targets, achieving state-of-the-art performance with more than 5 times faster inference compared to previous methods.
Key insights
- Transformers' self-attention mechanism effectively captures long-range dependencies, improving gaze target prediction, especially for distant targets where previous methods struggled.
- The iterative refinement within the multi-layer decoder progressively enhances predictions by modeling relationships between human-gaze-target instances and the overall scene context.
- Reframing human-gaze-target detection as a set-based prediction problem enables parallel processing of all instances, significantly boosting efficiency for multi-person scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.