Tracking Objects Through Assembly and Decomposition
Paper • 2611.26269 • Published
OmniST is a benchmark for tracking objects through both decomposition and assembly transformations, introduced in Tracking Objects Through Assembly and Decomposition.
OmniST is a dataset for evaluating object tracking through both decomposition and assembly transformations. It contains:
JPEGImages/ contains the video frames, one folder per video (<video>/00000.jpg, 00001.jpg, ...)Annotations/ contains the object and transformation annotations used for evaluation (<video>_annotations.json and <video>_transformations.json)ImageSets/val.txt lists the videos in the evaluation splitVideos/ contains the source videos as .mp4 files at 15 fps<video>_annotations.json)
Object masks, keyed by frame index and then by object ID:
{
"0": {
"0": {"object_name": "triscuit box", "counts": "<RLE>", "size": [576, 1024]},
"1": {"object_name": "bowl", "counts": "<RLE>", "size": [576, 1024]}
},
"1": { "...": "..." }
}
<video>_transformations.json)
A list of transformation events, for example:
{
"id": 0,
"start_frame": 45,
"end_frame": 65,
"type": "decomposition",
"start_state_objects": [{"id": 0, "name": "triscuit box"}],
"end_state_objects": [{"id": 0, "name": "triscuit box"}, {"id": 3, "name": "bag of triscuits"}],
"description": "remove,take_out"
}
type is "decomposition" or "composition" (assembly).ids refer to the object IDs in the matching <video>_annotations.json.Please refer to the code repository on how to evaluate OmniST.
Coming soon!