DynPose-100K is a large-scale dataset released by NVIDIA in 2025, containing 100,000 diverse dynamic videos with precise camera annotations. The associated research paper is「Dynamic Camera Poses and Where to Find Them」, aimed at addressing the challenge of accurately estimating camera intrinsics and poses in dynamic videos. By filtering videos suitable for camera estimation and improving dynamic video camera estimation algorithms, it provides high-quality data support for computer vision research.

This dataset is curated from 3.2 million candidate videos and includes video clips with rich dynamic content. All videos are annotated with world-to-camera poses and camera intrinsics.

Dataset Composition

The dataset mainly consists of the following components and fields:

• Video data: Approximately 100,000 MP4 format videos with a resolution of 1280×720 (720p), totaling about 200 GB.
• Frame data: JPG format frames extracted from the videos at a frame rate of 12 fps, totaling about 400 GB.
• Camera annotation data: PKL format files containing the following fields:
• poses: Camera pose matrices (N',3,4)
• intrinsics: Camera intrinsic matrices (3,3)
• frame_idxs: Corresponding frame index lists
• mean_reproj_error: Average reprojection error
• num_points: Number of reprojected points
• num_frames: Total number of video frames
• Metadata files:
• metadata.csv: Contains video IDs, URLs, timestamps, descriptions, etc., used for downloading videos and extracting frames.
• uid_mapping.csv: UID mapping file.
• Lightspeed benchmark set: Contains short video clips with ground-truth camera poses, used to validate the annotation method of DynPose-100K, totaling about 8.1 GB.

Citation

@inproceedings{rockwell2025dynpose,
    author = {Rockwell, Chris and Tung, Joseph and Lin, Tsung-Yi and Liu, Ming-Yu and Fouhey, David F. and Lin, Chen-Hsuan},
    title = {Dynamic Camera Poses and Where to Find Them},
    booktitle = {CVPR},
    year = 2025
}