Procedural Human Action Videos

by No License

Procedural Human Action Videos

This dataset contains about 40,000 videos for human action recognition that had been generated using a 3D game engine. The dataset contains about 6 million frames which can be used to train and evaluate models not only action recognition but also models for depth map estimation, optical flow, instance segmentation, semantic segmentation, 3D and 2D pose estimation, and attribute learning. (Cesar Roberto de Souza)

Dataset Attributes