For convenience, we provide precomputed detections for the training
, validation
, and testing
splits.
The file structure of the detections is as follows:
- detections
- training <-- 65 logs
- 0ef28d5c-ae34-370b-99e7-6709e1c4b929 <-- Log id
- per_sweep_annotations_amodal <-- for visualization
- tracked_object_labels_315969338019829000.json <-- Sweep detections
- validation <-- 24 logs
- testing <-- 24 logs
The structure shown above applies to both validation and testing as well.
The schema of each .json
file is defined below
{
"center": { <-- Center of the object cuboid, given as x, y, z
"x": 2.709212303161621,
"y": 3.548607349395752,
"z": 0.5272418856620789
},
"height": 1.771566390991211, <-- Object height
"label_class": "VEHICLE", <-- Argmax label from detector
"length": 4.258421897888184, <-- Object length
"occlusion": 0, <-- Not used
"rotation": { <-- Object pose represented as a quaternion
"w": -0.9999209457658587,
"x": 0.0,
"y": 0.0,
"z": 0.012573870474545222
},
"score": 0.9322837591171265, <-- Confidence for the corresponding `label_class`
"timestamp": 315972349019942000, <-- Corresponding LiDAR sweep timestamp
"track_label_uuid": null, <-- Not used
"tracked": true, <-- Not used
"width": 1.8624107837677002 <-- Object width
},
Note: Please refer to object_label_record.py
in argoverse-api
for loading information and usage.
Note: All of the logs found in the four parts of the training set are found in the training
folder.