trajgazemerge

In the Eye of MLLM: Benchmarking Egocentric Video Intent Understanding with Gaze-Guided Prompting

Taiying Peng¹, Jiacheng Hua², Miao Liu^2†, Feng Lu^1†
¹State Key Laboratory of VR Technology and Systems, School of CSE, Beihang University
²College of AI, Tsinghua University
NeurIPS D&B 2025

---

TrajGazeMerge — gaze-trajectory token selection for VisionZip (this fork)

This fork extends EgoGazeVQA with TrajGazeMerge: a study of gaze/hand-trajectory-guided visual-token selection for Qwen2.5-VL-7B under a fixed 10% token budget (VisionZip pruning). The headline result is M1 (VZ-complement, 63.01%) — selecting the gaze/hand-relevant tokens VisionZip's content attention discarded, added as a disjoint complement to its 7% content set.

• Best model — M1 (VZ-complement top-k, 63.01%): MODEL_M1_VZ_COMPLEMENT.md
• Full experiment grid & conclusions: TRAINING_RUNS.md
• Scanpath side-channel (63.01 tie, gated off): MODEL_SCANPATH_OURS.md
• Code: TrajGazeMerge/ (models, training, data, eval) · TrajGaze_v2/ (frozen TAS Stage-1 encoder) · scripts/ (launch scripts)

Protocol: egtea 2-way eval (n=1011), gaze-overlay frames, 3-epoch LoRA (eff-batch 8, early-stop), --merge-ratio 0.9. Model weights and raw datasets are not included in this repo.

---

Overview

This repository provides the official code for EgoGazeVQA, a benchmark for evaluating multimodal large language models (MLLMs) on egocentric video understanding tasks with gaze guidance.

Code Purpose:

• Generate gaze-guided QA pairs from egocentric videos (spatial, temporal, causal intent questions)
• Evaluate MLLMs with three gaze-guided prompting strategies (textual, visual marks, salience maps)
• Calculate and analyze model performance on intent understanding tasks

<p align="center"> <img src="assets/draw_intro.png" alt="EgoGazeVQA Overview" width="100%"> </p>

Timeline

• [2025-10] Code and dataset publicly released
• [2025-09] Paper accepted at NeurIPS 2025 D&B Track

TODO

• Add fine-tuning scripts for LoRA adaptation

Repository Structure

EgoGazeVQA/
├── generate_tool/
│   ├── auto.sh
│   ├── spatial.py
│   ├── temporal.py
│   ├── causal.py
│   └── create_datasets.py
├── test_tool/
│   ├── qwenvl_test/
│   │   ├── test_wo.py
│   │   ├── test_gaze.py
│   │   ├── test_mark.py
│   │   └── test_saliencemap.py
│   ├── prompt_gazees/
│   ├── multiframes/
│   ├── gaze_trajectory.py
│   └── caculate.py

Installation

git clone https://github.com/taiyi98/EgoGazeVQA.git
cd EgoGazeVQA

conda create -n egogazevqa python=3.10
conda activate egogazevqa

pip install -r requirements.txt

Dataset

Download from 🤗 Hugging Face:

huggingface-cli download taiyi98/EgoGazeVQA --repo-type dataset --local-dir ./data

Usage

Generate QA Pairs

cd generate_tool

# Generate for specific video
python spatial.py --video_id <VIDEO_ID> --target_index <INDEX>
python temporal.py --video_id <VIDEO_ID> --target_index <INDEX>
python causal.py --video_id <VIDEO_ID> --target_index <INDEX>

# Batch processing
bash auto.sh

Evaluate Models

cd test_tool/qwenvl_test

python test_wo.py              # Baseline (no gaze)
python test_gaze.py            # Textual gaze prompt
python test_mark.py            # Visual gaze marks
python test_saliencemap.py     # Gaze salience maps

Calculate Results

cd test_tool
python caculate.py --result_file <RESULT_CSV_PATH>

Citation

@misc{peng2025eyemllmbenchmarkingegocentric,
    title={In the Eye of MLLM: Benchmarking Egocentric Video Intent Understanding with Gaze-Guided Prompting}, 
    author={Taiying Peng and Jiacheng Hua and Miao Liu and Feng Lu},
    year={2025},
    eprint={2509.07447},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2509.07447}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.