LaMOT: Language-Guided Multi-Object Tracking
Yunhao Li, Xiaoqiong Liu, Luke Liu, Heng Fan, Libo Zhang
Figure: We introduce a large-scale benchmark LaMOT, which comprises 1,660 sequences, 1.67M frames, and over 18.9K trajectories. LaMOT contains videos from 4 datasets, including MOT17, TAO, SportsMOT, and VisDrone2019. It encompasses five different video scenarios including surveillance, autonomous driving, sports broadcasting, drone, and daily life. To our knowledge, LaMOT is the first benchmark dedicated to Language-Guided MOT.
We provide the annotations in the following manner:
annotations
├── train
| └──MOT17
| └──MOT17-02-FRCNN_01.json
| ...
| └──VisDrone
| └──uav0000013_00000_v_01.json
| └──uav0000013_00000_v_02.json
| ...
| └──SportsMOT
| └──v_00HRwkvvjtQ_c001_01.json
| └──v_00HRwkvvjtQ_c001_02.json
| ...
| └──TAO
| └──ArgoVerse
| └──00c561b9-2057-358d-82c6-5b06d76cebcf_01.json
| ...
| └──LaSOT
| └──airplane-3_01.json
| └──basketball-14_01.json
| ...
| ...
├── val
| ...
In these .json files, the last two digits represent the identifier number, while the preceding name corresponds to the sequence name.
In the .json files, we provide the language descriptions used and their corresponding track_ids. We note that we use the original track_ids provided by their source datasets. Specifically, we organize the annotations as follows:
{
"targets": {
'1': [1, 2, 3], # frame_id: [track_ids]
'2': [1, 2, 4],
...
},
"language": "a person walking on the street"
}For additional annotations of LaMOT, please refer to the annotations folder. For image sequences and bounding box annotations, please follow the download guidelines of MOT17, TAO, VisDrone2019, and SportsMOT.
We do not further partition the categories into base and novel classes. Therefore, the evaluation of LaMOT is the same as standard MOT datasets. We use the TrackEval Evaluation Toolkit.