Datasets:

danaleee
/

StreamGaze

language:
  - en
license: cc-by-4.0
task_categories:
  - video-text-to-text
tags:
  - gaze
  - multimodal
  - video-understanding
  - streaming-video
  - temporal-reasoning
  - proactive-understanding
  - egocentric-vision
  - visual-question-answering

StreamGaze Dataset

Paper | Project Page | Code

StreamGaze is a comprehensive streaming video benchmark for evaluating MLLMs on gaze-based QA tasks across past, present, and future contexts.

📁 Dataset Structure

streamgaze/
├── metadata/
│   ├── egtea.csv              # EGTEA fixation metadata
│   ├── egoexolearn.csv        # EgoExoLearn fixation metadata
│   └── holoassist.csv         # HoloAssist fixation metadata
│
├── qa/
│   ├── past_gaze_sequence_matching.json
│   ├── past_non_fixated_object_identification.json
│   ├── past_object_transition_prediction.json
│   ├── past_scene_recall.json
│   ├── present_future_action_prediction.json
│   ├── present_object_attribute_recognition.json
│   ├── present_object_identification_easy.json
│   ├── present_object_identification_hard.json
│   ├── proactive_gaze_triggered_alert.json
│   └── proactive_object_appearance_alert.json
│
└── videos/
    ├── videos_egtea_original.tar.gz         # EGTEA original videos
    ├── videos_egtea_viz.tar.gz              # EGTEA with gaze visualization
    ├── videos_egoexolearn_original.tar.gz   # EgoExoLearn original videos
    ├── videos_egoexolearn_viz.tar.gz        # EgoExoLearn with gaze visualization
    ├── videos_holoassist_original.tar.gz    # HoloAssist original videos
    └── videos_holoassist_viz.tar.gz         # HoloAssist with gaze visualization

🎯 Task Categories

Past (Historical Context)

Gaze Sequence Matching: Match gaze patterns to action sequences
Non-Fixated Object Identification: Identify objects outside gaze
Object Transition Prediction: Predict object state changes
Scene Recall: Recall scene details from memory

Present (Current Context)

Object Identification (Easy/Hard): Identify objects in/outside FOV
Object Attribute Recognition: Recognize object attributes
Future Action Prediction: Predict upcoming actions

Proactive (Future-Oriented)

Gaze-Triggered Alert: Alert based on gaze patterns
Object Appearance Alert: Alert on object appearance

🚀 Quick Start

Data Preparation

Download our dataset from HuggingFace and extract videos. The dataset should be located as below (note: the dataset itself is in danaleee/StreamGaze):

StreamGaze/
├── dataset/
│   ├── videos/
│   │   ├── original_video/        # Original egocentric videos
│   │   └── gaze_viz_video/        # Videos with gaze overlay
│   └── qa/
│       ├── past_*.json             # Past task QA pairs
│       ├── present_*.json          # Present task QA pairs
│       └── proactive_*.json        # Proactive task QA pairs

Extract Videos

# Extract EGTEA videos
tar -xzf videos_egtea_original.tar.gz -C videos/egtea/original/
tar -xzf videos_egtea_viz.tar.gz -C videos/egtea/viz/

# Extract EgoExoLearn videos
tar -xzf videos_egoexolearn_original.tar.gz -C videos/egoexolearn/original/
tar -xzf videos_egoexolearn_viz.tar.gz -C videos/egoexolearn/viz/

# Extract HoloAssist videos
tar -xzf videos_holoassist_original.tar.gz -C videos/holoassist/original/
tar -xzf videos_holoassist_viz.tar.gz -C videos/holoassist/viz/

Running Evaluation

Quick evaluation on existing models:

# Evaluate ViSpeak (without gaze visualization)
bash scripts/vispeak.sh

# Evaluate ViSpeak (with gaze visualization)
bash scripts/vispeak.sh --use_gaze_instruction

# Evaluate GPT-4o
bash scripts/gpt4o.sh --use_gaze_instruction

# Evaluate Qwen2.5-VL
bash scripts/qwen25vl.sh --use_gaze_instruction

Results will be automatically computed and saved to:

results/
├── ModelName/
│   ├── results/              # Without gaze visualization
│   │   ├── *_output.json
│   │   └── evaluation_summary.json
│   └── results_viz/          # With gaze visualization
│       ├── *_output.json
│       └── evaluation_summary.json

🔑 Metadata Format

Each metadata CSV contains:

video_source: Video identifier
fixation_id: Fixation segment ID
start_time_seconds / end_time_seconds: Temporal boundaries
center_x / center_y: Gaze center coordinates (normalized)
representative_object: Primary object at gaze point
other_objects_in_cropped_area: Objects within FOV
other_objects_outside_fov: Objects outside FOV
scene_caption: Scene description
action_caption: Action description

📝 QA Format

Each QA JSON file contains:

  {
    "response_time": "[00:08 - 09:19]",
    "questions": [
      {
        "question": "Among {milk, spoon, pan, phone}, which did the user never gaze at?",
        "time_stamp": "03:14",
        "answer": "A",
        "options": [
          "A. milk",
          "B. spoon",
          "C. pan",
          "D. phone"
        ],
      }
    ],
    "video_path": "OP01-R03-BaconAndEggs.mp4"
  }

📄 License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

See https://creativecommons.org/licenses/by/4.0/

🔗 Links

Paper: https://huggingface.co/papers/2512.01707
Evaluation code: https://github.com/daeunni/StreamGaze
Project page: https://streamgaze.github.io/

📧 Contact

For questions or issues, please contact: daeun@cs.unc.edu