This dataset was produced as part of our FOUND paper. The code there shows how to load and manipulate this dataset.
The dataset contains a folder for each multiview scan, with the following items:
Folder/file | Data type | Description |
---|---|---|
rgb |
Folder, .jpeg |
Captured images |
norm_gt |
Folder, .png |
Ground truth normals (from rendering mesh.obj |
norm_found |
Folder, .png |
Surface normals as predicted in FOUND |
norm_found_unc |
Folder, .png |
Surface normal uncertainty as predicted in FOUND |
normals_colmap |
Folder, .png |
Surface normals reconstructed via COLMAP |
mesh.obj |
.obj |
Ground truth 3D scan |
colmap_mesh.obj |
.obj |
3D scan from COLMAP |
keypoints.json |
.json |
2D keypoints as predicted in FOUND |
colmap.json |
.json |
Camera calibration from COLMAP |
These are predicted keypoints from a network trained on synthetic data, so their accuracy cannot be guaranteed.
{
"kp_labels": ["big toe", "2nd toe", ...], # N keypoint names
"annotations":
{
<IMG_PATH> : {
"kps": [[0., 0., 0.], ...], # [N x 2] Keypoints
"vis": [1.0, ...], # N visibility likelihoods (0 - 1)
"variance": [[1.0, 1.0], ...] # [N x 2] predicted variance of keypoint positions
}
...
}
}
The data here is taken from a COLMAP reconstruction, which has been scaled to match our ground truth meshes.
{
"camera":
{
"width": 480, # image width, pixels
"height": 640, # image height, pixels
"f": 500, # focal length, pixels
"cx": 240, # optical centre X, pixels
"cy": 320, # optical centre Y, pixels
"k": 0.001, # skew
},
"images":
[
{
"image_id": 1, # Used within COLMAP
"pth": "IMG_6268.jpeg", # filepath of image within rgb folder
"R": [[...], ...], # [3 x 3] Camera rotation matrix
"C": [...], # 3 world camera position
"T": [...], # 3 world camera position in 'R' reference frame
},
...
]
}
All matrices are given in COLMAP's coordinate system.