ITTO: Instruction-Tuned Tracking Objects Benchmark

ITTO is a real-world video benchmark for evaluating point tracking models under challenging motion and occlusion conditions. It highlights failure modes common in real-world settings but absent in many prior benchmarks, such as long-range motion, object reappearances, and complex scene dynamics.

📁 Contents

• High-quality point annotations on videos from MOSE, L-VOS, and Ego4D
• Split metadata and annotations
• Scripts to assist downloading Ego4D clips (requires license)

⚠️ Ego4D videos are not included. Only metadata and identifiers are provided.

📊 Statistics

• 72 real-world videos
• 1,373 annotated point tracks
• 11,449 total video frames
• Avg. track duration: 221.6 frames
• Avg. objects per video: 10.9
• Occlusion rate: 58.1%
• Reappearances per track: 5.86

🧪 Evaluation

Tracks are evaluated along key axes:

• Occlusion recovery and re-identification
• Motion complexity and long-term memory degradation
• Performance under dense scenes

📜 License

This benchmark is released under CC BY-NC-SA 4.0 due to MOSE.

• MOSE: CC BY-NC-SA 4.0
• L-VOS: BSD 3-Clause
• Ego4D: Metadata only – raw videos must be downloaded separately via ego4d-data.org

📦 Usage

from datasets import load_dataset

ds = load_dataset("your-username/itto", split="validation")