To download full WildRGB-D Dataset, it totally requires approximately 3.37T disk space to store zip packages, and approximately 4T to store all data.

To download all categories, execute python download.py --cat all.

To download specific one category, execute python download.py --cat <category_name>.

You could check all category names in the download scripts.

WildRGB-D
    ├── <category_name>               
    │   ├── scenes
    │   │   ├── scenes_<scene_id>
    │   │   │   ├── rgb
    │   │   │   │   ├── <frame_id>.png
    │   │   │   │   |
    │   │   │   ├── depth
    │   │   │   │   ├── <frame_id>.png
    │   │   │   │   |
    │   │   │   ├── masks
    │   │   │   │   ├── <frame_id>.png
    │   │   │   │   |
    │   │   │   ├── metadata
    │   │   │   ├── cam_poses.txt
    │	├── types.json
    │	├── nvs_list.json
    │	├── camera_eval_list.json

1. <category_name>/scenes/scenes_<scene_id>/depth/: We store depths in the depth scale of 1000. That is, when we load depth image and divide by 1000, we could get depth in meters.
2. <category_name>/scenes/scenes_<scene_id>/metadata: It stores the camera intrinsics including image width, height and K.
3. <category_name>/scenes/scenes_<scene_id>/cam_poses.txt: It stores the camera extrinsics. For every line, we list the <frame_id> first, then following the flatten 4x4 extrinsic matrix. Our camera extrinsics follows OpenCV convention, and it's camera to world matrix.
4. <category_name>/types.json: It stores the video type of every scene in <category_name>/scenes/. It includes single object video marked in "single", multi-object video marked in "multi" and hand-object video marked in "hand".
5. <category_name>/nvs_list.json: It stores the training and validation split we use in our Novel View Synthesis Task. For Single-Scene NVS, we only test on val split. For Cross-Scene NVS, we pre-train on train split and test on val split.
6. <category_name>/camera_eval_list.json: It stores the training and validation split we use in our Camera Pose Evaluation Task.

Our WildRGB-D Dataset provides point cloud annotations. Please refer to wildrgbd_generate_point_cloud.py.

If you have any problems when downloading and using WildRGB-D Dataset, please contact Hongchi Xia by email.