MolmoPoint-TrackAny: Point Tracking Annotations for Natural Videos

AllenAI's MolmoPoint-TrackAny dataset provides 17,437 unique point tracking annotations across 13,050 natural videos sourced from YouTube and MammalNet. Each annotation includes an object description, per-frame point trajectories, and video metadata, with two temporal sampling variants (1 and 2 FPS) concatenated into a single training split. The dataset was last updated on March 17, 2026.

Use Cases

• Train point tracking models based on per-frame point trajectory annotations.
• Benchmark video object tracking performance based on natural video examples.
• Develop models for object localization from textual expressions based on the provided object descriptions.
• Analyze temporal consistency in tracking across different frame rates based on the 1 and 2 FPS variants.

Strengths

• Contains 17,437 unique annotations, providing a substantial number of examples.
• Annotations are sourced from two distinct platforms, YouTube and MammalNet.
• Includes two temporal sampling variants (1 and 2 FPS) for each annotation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect geographic or content bias inherent to the YouTube and MammalNet sources.

Provenance

Source

AllenAI, sourced from YouTube and MammalNet.

Collection Method

Annotations created for point tracking in natural videos.

Time Range

null

Freshness

Last updated 2026-03-17 16:51:36; freshness should be verified.

Geography

null