The dataset is organized as follows:
.
├── images # contains all images
├── train
├── val
├── posetrack_data # contains annotations for pose reid tracking
├── train
├── 000001_bonn_train.json
├── ...
├── val
├── ...
├── posetrack_mot # contains annotations for multi-object tracking
├── mot
├── train
├── 000001_bonn_train
├── image_info.json
├── gt
├── gt.txt # ground truth annotations in mot format
├── gt_kpts.txt # ground truth poses for each frame
├── ...
├── val
├── posetrack_person_search # person search annotations
├── query.json
├── train.json
├── val.json
We adopt the same format proposed in PoseTrack18. Please note that keypoints for left_ear and right_ear are just placeholders for compatibility with the MSCOCO dataset.
A sample annotation file in posetrack_data has the following format:
{
"images": [
{
"is_labeled": false, # true, if a frame contains annotated keypoints
"nframes": 76
"frame_id": 10000010000,
"id": 10000010000,
"vid_id": 000001,
"file_name": images/train/000001_bonn_train/000000.jpg,
"has_labeled_person": false, # true, if a frame contaisn annotated bounding boxes (with or without keypoints)
"ignore_regions_y": [
[y11, y12, y13, ..., y1n],
[y21, y22, y23, ..., y2n]
],
"ignore_regions_x": [
[x11, x12, x13, ..., x1n],
[x21, x22, x23, ..., x2n]
]
}
],
"annotations": [
{
"bbox": [x1, y1, w, h],
"bbox_head": [x1, y1, w, h],
"category_id": 1,
"image_id": 10000010000,
"id": 1000001000000,
"keypoints": [x1, y1, vis1, ..., x17, y17, vis17],
"person_id": 1024,
"track_id": 0
}
],
"categories": [
{
"supercategory": "person",
"id": 1,
"name": "person",
"keypoints": [
"nose",
"head_bottom",
"head_top",
"left_ear",
"right_ear",
"left_shoulder",
"right_shoulder",
"left_elbow",
"right_elbow",
"left_wrist",
"right_wrist",
"left_hip",
"right_hip",
"left_knee",
"right_knee",
"left_ankle",
"right_ankle"
],
"skeleton": [
[
16,
14
],
[
14,
12
],
[
17,
15
],
[
15,
13
],
[
12,
13
],
[
6,
12
],
[
7,
13
],
[
6,
7
],
[
6,
8
],
[
7,
9
],
[
8,
10
],
[
9,
11
],
[
2,
3
],
[
1,
2
],
[
1,
3
],
[
2,
4
],
[
3,
5
],
[
4,
6
],
[
5,
7
]
]
}
]
}
Similarly, submission files should have the following format:
{
"images": [
{
"file_name": "images/train/000001_bonn_train/000000.jpg",
"id": 10000010000,
"frame_id": 10000010000
},
],
"annotations": [
{
{
"bbox": [x1, y1, w, h],
"image_id": 10000010000,
"keypoints": [x1, y1, vis1, ..., x17, y17, vis17],
"scores": [s1, ..., s17],
"person_id": 1024,
"track_id": 0
}
}
]
}
Note, that the vis1 to vis17 are ignored during evaluation.
In MOT, we provide a slightly different format, compared to related MOT datasets such as MOT17. In particular, we replace seqinfo.ini by image_info.json with the following format:
[
{
"file_name": "images/train/000001_bonn_train/000000.jpg",
"id": 10000010000,
"frame_id": 10000010000,
"vid_id": "000001",
"frame_index": 1, # the frame index refered to in gt.txt
"ignore_regions_x": [
[x1, x2, ..., xn]
],
"ignore_regions_y": [
[y1, y2, ..., yn]
]
},
...
]
Each gt.txt follows the same format proposed by the MOT17 dataset, where <x>, <y>, <z> are set to -1 and <conf> is set to 1:
<frame>, <id>, <bb_left>, <bb_top>, <bb_width>, <bb_height>, <conf>, <x>, <y>, <z>
posetrack_person_search/train.json, posetrack_person_search/posetrack_person_search/val.json and posetrack_person_search/posetrack_person_search/query.json follow a very similar structure as the annotations for pose reid tracking.
The training set contains 5474 unique person ids.