OmniPro

A Comprehensive Benchmark for Omni-Proactive Streaming Video Understanding

OmniPro evaluates multimodal models on their ability to proactively interact with streaming video — detecting events, monitoring states, counting objects, and providing timely narrations without explicit user queries at each moment.

📋 Overview

OmniPro consists of 9 evaluation tasks across two modes:

Tasks

Evaluation Modes

• Probe Mode: Model receives a video clip up to time t and answers whether an event has occurred. Tests temporal awareness and content understanding.
• Online Mode: Model processes video frame-by-frame in real-time and autonomously decides when to speak and what to say.

🏗️ Project Structure

OmniPro/
├── models/                 # Model adapters (probe + streaming)
│   ├── base.py            # BaseModel abstract class
│   ├── streaming_base.py  # StreamingModel abstract class
│   ├── qwen3_vl.py        # Qwen3-VL probe adapter
│   ├── qwen2_5_omni.py    # Qwen2.5-Omni (audio+visual)
│   ├── qwen3_omni.py      # Qwen3-Omni
│   ├── internvl3.py       # InternVL3.5-8B
│   ├── phi4_multimodal.py # Phi-4-multimodal
│   ├── video_salmonn2.py  # Video-SALMONN2+
│   ├── videollama2_av.py  # VideoLLaMA2.1-7B-AV
│   ├── livestar.py        # LiveStar-8B (online)
│   ├── livestar_probe.py  # LiveStar-8B (probe)
│   ├── mmduet2.py         # MMDuet2 (online)
│   ├── mmduet2_probe.py   # MMDuet2 (probe)
│   ├── minicpm_o.py       # MiniCPM-o 4.5 (online)
│   ├── minicpm_o_probe.py # MiniCPM-o 4.5 (probe)
│   ├── gemini.py          # Gemini-3-Flash (API)
│   └── ...
├── evaluators/             # Evaluation engines
│   ├── probe_evaluator.py # GT-probe evaluation logic
│   └── online_evaluator.py# Frame-by-frame streaming evaluation
├── metrics/                # Scoring and metrics
│   ├── probe/             # Probe metrics (paired accuracy, F1)
│   └── online/            # Online metrics (time F1, content accuracy)
├── utils/                  # Utilities
│   ├── prompts.py         # Task-specific prompt templates
│   ├── video.py           # Video splitting/processing
│   ├── io.py              # Data I/O
│   └── online_parser.py   # Response parsing
├── scripts/                # Run scripts (one-click evaluation)
│   ├── run_probe.py       # Probe evaluation entry point
│   ├── run_online.py      # Online evaluation entry point
│   ├── compute_metrics.py # Compute probe metrics
│   ├── compute_online_metrics.py # Compute online metrics
│   ├── run_probe_*.sh     # Per-model probe scripts
│   └── run_online_*.sh    # Per-model online scripts
├── data/
│   └── benchmark.json     # Benchmark annotations (2700 samples)
├── third_party/            # Third-party model code (see README inside)
├── visualization/          # Demo and visualization
└── requirements.txt        # Python dependencies

🚀 Quick Start

1. Environment Setup

# Clone the repository
git clone https://github.com/RuixiangZhao/OmniPro
cd OmniPro

# Install base dependencies
pip install -r requirements.txt

# Clone third-party model repos (see third_party/README.md)

2. Data Preparation

Clone the data repository (videos + metadata) into the data/ directory:

# Clone the benchmark data (videos, annotations)
cd data
git clone https://huggingface.co/datasets/RuixiangZhao/OmniPro .
cd ..

The data/ directory should contain:

• benchmark.json — Benchmark annotations (2700 samples, 9 tasks × 300)
• raw_videos/ — Source video files referenced by benchmark.json

3. Run Probe Evaluation

# Example: Qwen3-VL-8B
bash scripts/run_probe_qwen3_vl.sh

# Quick smoke test (2 samples/task)
LIMIT=2 bash scripts/run_probe_qwen3_vl.sh

# Custom: specific tasks, limited samples
python scripts/run_probe.py \
    --model qwen3-vl \
    --model_path /path/to/model \
    --tasks instant_event_alert,event_narration \
    --limit 50 \
    --num_gpus 8 \
    --output_dir results/probe/Qwen3-VL-8B/

4. Run Online Evaluation

# Example: MiniCPM-o 4.5
bash scripts/run_online_minicpmo.sh

# Quick test
LIMIT=4 bash scripts/run_online_minicpmo.sh

5. Compute Metrics

# Probe metrics
python scripts/compute_metrics.py --pred_dir results/probe/Qwen3-VL-8B/ --tolerance 3,5

# Online metrics
python scripts/compute_online_metrics.py --pred_dir results/online/MiniCPM-o-4.5-Duplex/ --tolerance 3

📊 Supported Models

Probe Mode

Online Mode

📐 Metrics

Probe Mode

• Paired Accuracy: Both pre-probe and post-probe must be correct
• Content F1: F1 score across all probe points
• Pre/Post Accuracy: Separate accuracy for before/after trigger

Online Mode

• Time F1: Precision × Recall of emit timestamps within tolerance window
• Content Accuracy: Correctness of emitted content (parsed or GPT-judged)
• Joint F1: Combined timing + content score

🔧 Adding a New Model

Probe Mode

1. Create models/your_model.py inheriting from BaseModel
2. Implement name() and generate(instruction, video_path)
3. Register in scripts/run_probe.py (choices + build_model)
4. Create scripts/run_probe_your_model.sh

Online Mode

1. Create models/your_model.py inheriting from StreamingModel
2. Implement begin(), observe(), end()
3. Register in scripts/run_online.py
4. Create scripts/run_online_your_model.sh

📝 Citation

@article{omnipro2026,
  title={OmniPro: A Comprehensive Benchmark for Omni-Proactive Streaming Video Understanding},
  author={Zhao, Ruixiang and Yang, Jie and Xin, Zijie and Wang, Tianyi and Rao, Fengyun and LYU, Jing and Li, Xirong},
  journal={arXiv preprint arXiv:2605.18577},
  year={2026}
}

📄 License

This project is released under the MIT License.