Sign in to view source links and access this dataset
Description
Molmo2-VideoTrackEval is an evaluation benchmark for video point tracking containing human-annotated ground truth expressions. It includes segmentation masks for evaluating whether predicted points fall within the correct object regions across five categories: animal, dance, sports, person, and misc. The dataset is part of the Molmo2 collection and was created by AllenAI, with a last recorded update on December 16, 2025.
Use Cases
Benchmarking video point tracking models based on human-annotated ground truth expressions.
Evaluating segmentation mask accuracy for predicted object regions.
Comparing model performance across specific categories like sports or dance.
Training or fine-tuning video tracking algorithms using annotated video data.
Strengths
Contains human-annotated ground truth expressions, which suggests high-quality labels.
Includes segmentation masks for evaluating spatial accuracy of tracked points.
Provides a structured evaluation framework across five distinct categories.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
AllenAI
Collection Method
Human-annotated ground truth expressions.
Freshness
Last updated 2025-12-16 15:47:41; freshness should be verified.
License is unknown; users should verify terms of use before downloading.