Dynamic Replica
收藏资源简介:
该数据集是一个基准数据集,包含了在扫描环境中动画人类和动物的合成视频,旨在用于训练和评估立体深度估计中的时间一致性。数据集还包括了立体视图的地面真实深度图、光流、分割掩模以及相机参数,环境多样且相机基线随机化,以增强泛化能力。该数据集规模包含524个合成的人类和动物视频,其任务重点是研究从立体视频中估计视差的时间一致性。
This is a benchmark dataset featuring synthetic videos of animated humans and animals in scanned environments, designed for training and evaluating temporal consistency in stereo depth estimation. It includes ground-truth depth maps, optical flow, segmentation masks, and camera parameters for stereo views. To enhance generalization capability, the dataset incorporates diverse environments and randomized camera baselines, and contains 524 synthetic human and animal videos. The core focus of this dataset is investigating the temporal consistency of disparity estimation from stereo videos.




