Datasets with ground truth have driven many of the recent advances in computer vision. They allow evaluation and comparison so the field knows what works. They also provide training data to machine learning methods that are hungry for data. Creating good datasets that are valuable to the community and have a reasonable lifespan is hard work. Key issues are the quality and quantity of the data, how well that data addresses a specific problem in the field, and whether it is well curated with a good evaluation.
We have played central roles in many influential datasets and evaluations in the field including
- Middlebury flow dataset
- MPI-Sintel flow and related datasets
- KITTI datasets
- HumanEva for human pose estimation
- FAUST for 3D mesh registration
- JHMDB for action recognition
- MPI-I human pose dataset
Note that Perceiving Systems is involved in all three of the standard benchmarks in optical flow (Middlebury, KITTI and Sintel).
We are committed to releasing data whenever possible including
- Dyna: 40,000 4D human body scans
- Motion capture of extreme human poses