I have several questions about training and evaluation.
1. At github, there is a limitation of pretraining such as,
"Other tracking datasets (e.g., MOT20/) used for pretraining are forbidden."
Then, does it means that pretraining models with additional datasets except for multi-target tracking is acceptable?
2. there are the limited number of instances for single sequence. Therefore, is it acceptable to define maximum_instance_id=20 for instance assignment?
Thank you.
Posted by: icj_track @ June 24, 2022, 6:49 a.m.Thanks for your questions.
1. In this compition, we limit the use of TRACKING dataset to SportsMOT.
You can use datasets in other fields, like COCO for detection and market1501 for reid, to benefit the model. However, you are not allowed to use datasets of TRACKING like MOT15/17, DanceTrack, etc.
2. We have no explicit or mandatory rules about this. It's indeed your choice.
Nevertheless, I want to mention that we aim at tracking diverse sports scenes. Even if the trick works, this limitation kind of deviates from the goal of SportsMOT.
I appreciate your quick reply and it is very helpful.
Then, I have more questions about additional dataset.
1. Is it acceptable to use only visual information(e.q. image, clips) of tracking dataset(e.q. MOT17,20) without annotations?
2. Moreover, gathering additional unannotated sports videos from internet and pseudo-labelling them by employing the detector pre-trained on sportsmot can be applicable to train a model for competition?
Thank you!
Posted by: icj_track @ June 29, 2022, 5:53 a.m.I have more question about comparison between MultiSports and SportsMOT.
1. I am curious that SportsMOT's videoset is a subset of MultiSports' videoset or there exists un-overlapped videos in SportsMOT.
Thank you.
Posted by: icj_track @ June 30, 2022, 3:37 a.m.Yes. Videos used in SportsMOT are of a subset of source videos of MultiSports. Besides, clips in SportsMOT are manually selected to ensure there are no shot changes and to compose balanced categories, etc.
Posted by: XiaoyuZhao @ June 30, 2022, 4:09 a.m.I appreciate your quick reply and it is very helpful.
Could you give me a response for the former questions as below?
1. Is it acceptable to use only visual information(e.q. image, clips) of tracking dataset(e.q. MOT17,20) without annotations?
2. Moreover, gathering additional unannotated sports videos from internet and pseudo-labelling them by employing the detector pre-trained on sportsmot can be applicable to train a model for competition?
Thank you!
Posted by: icj_track @ June 30, 2022, 5:54 a.m.It's my pleasure to offer help :)
1. This is indeed your choice. Nevertheless, it should be mentioned that MOT datasets mainly focus on tracking pedestrians in crowded scenes. Please consider the priors in data before trying to utilize other datasets besides SportsMOT.
2. Like the former question, we do not want to restrict this and it's your choice.
IMPORTANT: If you want to use these kind of "tricks" in the final test submissions, please mention them honestly in the report.
Posted by: XiaoyuZhao @ June 30, 2022, 10:29 a.m.Thank you for your considerable responses.
I have one more question. Could I know the original source of sportsmot datasets? (e.q. parsing from other datasets or manually gathering from YouTube.)
Actually, I want to use other additional videos such as soccer-net(https://www.soccer-net.org/download) to train our model. But, I am curious that there are overlapped videos between sportsmot and the datasets.
Thank you!
Best regards.
video clips in SportsMOT are manually selected and checked. The source videos are a subset of the videos in MultiSports.
See "Data" section for more details.