Our dataset reads all matadata from a common json format. Each scene has a json which contains all the metadata required by the dataset. This includes a list of image files, a list of depth files, a list of poses, etc.
We provide a function to parse Scannet scenes into our format. Note that this does not duplicate or modify any of the original data, it just stores the required metadata (including paths) in a info.json file. To use with another dataset you need to generate these json files, and then you can use our dataset.
{'dataset': 'scannet',
'path': path,
'scene': scene,
'file_name_mesh_gt': '',
'instances': {}, # dict mapping instance id to label id
'frames': [{'file_name_image': '',
'file_name_depth': '',
'file_name_instance': '',
'intrinsics': intrinsics,
'pose': pose,
}
...
]
'file_name_seg_indices': '', # optional (used for labeling the tsdf from scannet labels)
'file_name_seg_groups': '', # optional (used for labeling the tsdf from scannet labels)
}