Detector issue
zhu011 opened this issue · comments
Hello authors! Really great job and inspires me a lot. One of the questions about the code of the article is that if the detection network is to be changed to a single-stage network similar to YOLO, what changes do you need to make to your code, whether you want to delete the loss of the first stage in your code, or what changes are needed elsewhere, I hope to get your answer
Hi, Thanks for your interest in this. There are two losses of detection that needs to be taken care of. First is the classification loss of the proposals and second is regression loss. I think to apply KCR to single stage, you need to apply KCR assignment rules to address the classification loss signal.
To make the regression projection easier, you can follow the formulation of oriented RCNN, using (x, y, w, h, alpha, beta). In that case, you can easily regress (x, y, w, h). Hopefully that helps.
about load_from = "/data/alan/q1/workspace/mmStorage/KCR/orcnn_dota1_noship/epoch_50.pth Can you provide this weight？
As for whether you use horizontal labeling to train the target dataset during the training process, then test whether you use rotational labeling to calculate the map of the target dataset, but why is the HRSC dataset test.json you give a horizontal label with four degrees of freedom?