omniproact-bench/OmniPro_subset

OmniPro Subset (Review Sample)

This is a lightweight subset of OmniPro for convenient reviewer inspection.

Contents

• 270 samples (30 per sub-task × 9 sub-tasks), randomly sampled from the full 2,700-sample benchmark
• 245 source videos corresponding to the sampled annotations
• benchmark_subset.json: all 270 evaluation samples with full annotations

Task Types

Task	Abbr.	Samples
Instant Event Alert	Event-Alert	30
Semantic Condition Alert	Cond.-Alert	30
Explicit Target Grounding	Target-Ground	30
Snapshot Counting	Snap.-Count	30
Cumulative Counting	Cum.-Count	30
Dedup Counting	Dedup.-Count	30
Realtime State Monitor	State-Monitor	30
Event Narration	Event-Narr.	30
Sequential Step Instruction	Step-Inst.	30

Sampling Method

Random sampling with seed=42, stratified by task type (30 samples per task), preserving the original distribution of audio dependencies and trigger types.

Relationship to Full Dataset

This subset is intended solely for quick quality inspection. The complete benchmark (2,700 samples, 1,262 videos, 32.6 GB) is available at:

👉 omniproact-bench/OmniPro

Evaluation code is available at: 🔗 OmniPro Evaluation Code

File Structure

.
├── README.md
├── benchmark_subset.json    # 270 evaluation samples
└── raw_videos/              # 245 source video files
    ├── {video_id_1}.mp4
    ├── {video_id_2}.mp4
    └── ...

License

CC BY-NC 4.0