CamVid-30K Redefines the Benchmark for Dynamic Scene Research
The dataset contains three major subsets such as VIPSeg and OpenVid, which completely record the four-dimensional elements such as camera position, object motion trajectory, etc., and the data scale reaches 30TB level. Its features are: 1) the precise camera parameters stored in COLMAP format support millimeter-level motion reconstruction; 2) the continuous interframe optical flow annotation realizes timing consistency verification; 3) the WebVid-10M subset provides diverse scene materials. Experimentally, it is demonstrated that the model trained with this dataset improves the PSNR metric by 12.7% in the NeRF motion reconstruction task, which has become a commonly used benchmarking platform for academic papers. The dataset follows the CC-BY-NC protocol and has been integrated into the PyTorch ecosystem.
This answer comes from the articleGenXD: open source framework for generating videos of arbitrary 3D and 4D scenesThe
