PointMotionBench: 3D Point Motion in Egocentric and Third-Person Video

A benchmark for evaluating 3D point motion in video, covering egocentric and third-person scenes across three source datasets. It contains 2,565 video clips from DAVIS and HOT3D datasets, each paired with per-object tracked surface points and a human-verified natural-language caption. The dataset was created by AllenAI and last updated on 2026-06-17.

Use Cases

• Benchmarking 3D point tracking algorithms based on the provided 2D and 3D tracked surface points.
• Training models for natural-language video captioning based on the human-verified captions.
• Evaluating scene understanding across egocentric and third-person viewpoints as described in the dataset overview.
• Developing multimodal models that link visual motion in video clips with textual descriptions.

Strengths

• Contains 2,565 video clips (90 from DAVIS, 2,475 from HOT3D).
• Provides both 2D and 3D tracked surface points per object.
• Includes human-verified natural-language captions for each sample.
• Covers two distinct scene types: third-person and egocentric.

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 bias inherent to the source datasets DAVIS and HOT3D.

Provenance

Source

AllenAI, aggregated from DAVIS and HOT3D datasets.

Collection Method

Likely compiled from existing video datasets with added annotations.

Freshness

Last updated 2026-06-17 03:06:16; freshness should be verified.