DynPose-100K: 100,000 Dynamic Camera Poses for Video Annotation

DynPose-100K provides 100,000 dynamic camera pose annotations, released in 2025 by researchers from NVIDIA, the University of Michigan, and NYU. The dataset focuses on camera trajectory estimation in video sequences and is accompanied by the Lightspeed benchmark for ground-truth validation.

Use Cases

• Training machine learning models for camera trajectory prediction using pose annotations
• Validating pose estimation algorithms against the ground-truth data in the Lightspeed benchmark
• Developing structure-from-motion (SfM) pipelines for dynamic video scenes

Strengths

• 100,000 pose records
• Includes Lightspeed benchmark for ground-truth validation
• Authored by researchers from NVIDIA, NYU, and University of Michigan

Limitations

• Specific column names and data formats are not documented in the summary
• License is categorized as 'other', requiring manual verification of usage rights
• Geographic scope is limited to US-based research collection

Provenance

Source

NVIDIA, University of Michigan, and New York University (Arxiv 2504.17788)

Collection Method

Annotated camera poses from video sequences

Time Range

2025

Freshness

Last updated May 2025 with the release of the Lightspeed benchmark.

Geography

United States