Another DeeperAction challenge "SportsMOT track" is the task tracking players. Videos of SportsMOT dataset probably include those of MultiSports dataset. Is there any problem for using SportsMOT dataset to track person boxes on MultiSports challenge?
Posted by: InwoongLee @ June 30, 2022, 1:04 a.m.No. This is not allowed.
Posted by: Judie1999 @ June 30, 2022, 5:37 a.m.Thank you for the reply.
I have another question. Is is possible to use another dataset such as coco, CrowdHuman, MOT17 datasets and etc. to track person boxes not SportsMOT?
No. You can only use the provided data from the MultiSports dataset and no extra data is allowed.
Posted by: Judie1999 @ June 30, 2022, 6:41 a.m.Sorry to bother you again. In https://github.com/MCG-NJU/MultiSports, they provide precomputed person boxes generated by the person detector of Faster R-CNN with a ResNeXt-101-FPN backbone. In my knowledge, MultiSports dataset has only actor boxes and labels not including other person. The reply "the MultiSports dataset and no extra data is allowed" means that we can't use any pretrained detector model from mot, coco, etc.? Should I detect actor person box and action labels only using MultiSports dataset without any pretrained detector from other datasets(coco, mot)? Is it impossible to use the pretrained model from other dataset?
Posted by: InwoongLee @ June 30, 2022, 7:01 a.m.Additionally, should I perform from scratch training only on MultiSports dataset? Is it impossible to use the pretrained weight from other datasets? I wonder "Only Multisports dataset" means from scratch training on MultiSports.
Posted by: InwoongLee @ June 30, 2022, 7:29 a.m.Hi! Sorry for the late reply. Last year we did not have the SportsMOT dataset so we did not have this problem. After discussion, we believe it is unfair to use boxes from the SportsMOT dataset, but other publicly available data should be allowed. We will send an email to clarify this. Thank you for your feedback!
Posted by: Judie1999 @ June 30, 2022, 7:52 a.m.