duanzhiihao
commented
3 years ago According to your source code, the anchor box is only vertical.
Yes, that's true.
How can the model learn the width, height and angle properly for rotating bounding box when there are no rotated anchors?
We didn't try rotated anchors. The model just miraculously learns how to predict the rotation from vertical anchors. Based on our experiments, the most critical thing is the massive training data from COCO using rotation augmentation. I guess the model somehow learns the pattern of different orientations given enough examples. It also learns the width/height invariant to the orientation. If you want to ask how that works in the internal CNN, that is beyond my knowledge. Rotated anchors might help, but it will also increase training and testing time complexity.
How do anchors work in your project?
They are the same as in standard object detection CNNs. We are predicting offsets from the anchors.
GivralNguyen
According to your source code, the anchor box is only vertical . Is this true, or am i missing something? How can the model learn the width, height and angle properly for rotating bounding box when there are no rotated anchors? How does anchors work in your project? Thanks in advance since i am not sure about this problem .