Can TSD head be used on top of RetinaNet?

[CodeJjang]

As I understood, the TSD fine tunes proposals from the RPNs existing in Faster RCNN, Mask RCNN, Cascada RCNN etc.
In RetinaNet we don't have region proposals but instead the head convolves the different levels of the FPN using anchors.
Theoretically, what if a certain spatial convolution location is good for the classification but a slightly offset one is better for regression, just like in the TSD case?
Wouldn't RetinaNet benefit from a TSD head as well?

[songguanglu]

Hi, this is a very meaningful question. In fact, TSD does help alleviate the misalignment between classification and localization in RetinaNet. In our private experiments, we have improved the performance of some anchor-free detectors by a large margin with the assistance of TSD.

[CodeJjang]

@songguanglu Do you have somewhere the implementation of TSD on RetinaNet you've done, for reference?

[songguanglu]

We have only performed the TSD variant on FCOS. If you are interested in the performance on RetinaNet, you can try to implement it based on this project. Any questions can be discussed here in time.

[BlueCCircles]

@songguanglu Thanks for your wonderful works on feature disentagle. I have the same question that dose the TSD supports anchor-based single stage detector ? It's a great news that TSD can improves the performance of FCOS, did you add it after FPN and the separatly heads used to get classification head and regression head, respectively?

[houlei123123]

您好，您是如何将TSD运用到FCOS中的呀，按照论文中提的，应该只能使用在two-stage中。