Reproduce “In the Eye of MLLM: Benchmarking Egocentric Video Intent Understanding with Gaze-Guided Prompting” 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
Taiying Peng, Jiacheng Hua, Miao Liu, Feng Lu
Researchers from Beihang and Tsinghua Universities introduced EgoGazeVQA, the first benchmark to integrate user gaze data for evaluating multimodal large language models' (MLLMs) understanding of human intent in egocentric videos. The study demonstrates that gaze-guided prompting consistently enhances MLLM performance, particularly in spatial reasoning, with LoRA fine-tuning on domain-specific data boosting a lighter MLLM's spatial reasoning accuracy from 40.4% to 67.7%.
Key insights
- Gaze information consistently improves MLLM performance in understanding human intent from egocentric videos, particularly for spatial and temporal reasoning.
- Larger MLLMs benefit more from gaze signals, suggesting that substantial model capacity is crucial for effective gaze interpretation.
- Gaze Salience Maps (GazeS) are most effective for spatial intent, while textual gaze prompts (GazeT) show greater gains for temporal intent understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.