Reproduce “Eyes on VLM: Benchmarking Gaze Following and Social Gaze Prediction in 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
Hengfei Wang, Anshul Gupta, Pierre Vuillecard, Jean-Marc Odobez
EyeVLM evaluates whether vision-language models can locate a person’s gaze target and predict social gaze relations. It asks models image questions conditioned on one or two head boxes, then compares prompts and task-specific fine-tuning with specialized gaze systems. On GazeFollow, fine-tuned Qwen3-VL-8B-Instruct reduces average L2 error to 0.110 from 0.215, but remains above MTGS-DINO-static’s 0.094.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.