2,000 synthetic video sequences featuring multiple interacting humans with dense ground truth for motion and segmentation. The dataset provides high-resolution frames designed to train models for complex optical flow estimation in multi-person scenarios.
Use Cases
- Train deep neural networks for optical flow using the flow field ground truth and RGB image sequences
- Perform instance-level motion segmentation using the provided human masks and motion vectors
- Benchmark multi-human tracking performance by leveraging the unique instance IDs across frames
Strengths
- 2,000+ synthetic video sequences with ground truth optical flow
- Includes per-pixel instance segmentation masks for all human subjects
- Provides depth maps and camera parameters for 3D motion analysis