ImageNet Video Visual Relationship Detection Dataset

Imagenet Vidvrd is a video dataset for visual relationship detection, derived from the ImageNet-VidVRD benchmark. The dataset was uploaded to HuggingFace by user 'shangxd' in January 2024. It contains video sequences annotated for object interactions and spatio-temporal relationships.

Use Cases

• Train object detection models to identify and localize entities like 'person' or 'vehicle' across video frames.
• Develop models for predicting predicate relationships, such as 'person riding bicycle', using annotated subject-object pairs.
• Benchmark spatio-temporal localization algorithms on annotated video segments with temporal boundaries.
• Perform video captioning or dense video description by leveraging the annotated visual relationships and their temporal evolution.

Strengths

• Based on the established ImageNet-VidVRD benchmark, providing a standard for evaluation.
• Focuses on spatio-temporal visual relationship detection, a complex and advanced computer vision task.

Limitations

• Specific row count, video duration, and class distribution details are unknown from the provided metadata.
• Potential for class imbalance common in relationship detection datasets, where certain predicates are rare.

Provenance

Source

Derived from the ImageNet-VidVRD benchmark.

Collection Method

Video sequences annotated with visual relationships (subject-predicate-object triples) and their spatio-temporal tubes.

Time Range

null

Freshness

Dataset was last updated on the platform in January 2024.

Geography

null