HRWSI stands for High-Resolution Web Stereo Image. It is a dataset for image depth estimation, consisting of approximately 21,000 high-resolution RGB-D images from the Internet. After removing the damaged ground truth data, 20,378 images are used for training and 400 images are used for verification. In addition, the dataset also provides sky segmentation masks, instance segmentation masks, and invalid pixel masks.

Compared with the ReDWeb dataset, the HRWSI dataset has the following three advantages: more training samples (20,378 vs. 3,600), higher resolution (833 × 1,251 vs. 408 × 465), and better handling of incorrect difference regions such as the sky.