Whether QDTrack supports changing detectors or using external detection results for tracking

[unihornWwan]

First of all, thank you for your outstanding contribution. For this paper, Faster R-CNN is used as the detector and RPN output is used to carry out quasi-dense sampling and similarity learning, indicating that RoI feature is needed in the tracking process. Does this mean that external detection results cannot be used for tracking in customized scenes，and we have to train the entire network from detectors to trackers? Thank you !

[pixeli99]

Yes, you're right. We need to retrain the entire network

[unihornWwan]

Yes, you're right. We need to retrain the entire network

Thank you for your reply. In addition, if I want to replace the detector, do I also have to use a two-stage detector with RPN or a single-stage detector with dense sampling like FCOS? Thank you!