Hi, nice to hear that you wanna try DAI-Net with YOLOv9. As I'm not very familiar with YOLOv9, I could only provide some general suggestions now:
I would suggest you start from establing the reflectance decoder, because it brings major improvements and, unlike redecomposition cohering loss, doesn't largely modify the processing procedure of the base detector.
You would need to choose a proper intermediate layer in the backbone to incorporate the decoder. For better efficiency, you could save the visualization results of the predicted R during training. If your implementation works well, the visualized R would look good in the very early training stage.
Hi, nice to hear that you wanna try DAI-Net with YOLOv9. As I'm not very familiar with YOLOv9, I could only provide some general suggestions now:
I would suggest you start from establing the reflectance decoder, because it brings major improvements and, unlike redecomposition cohering loss, doesn't largely modify the processing procedure of the base detector.
You would need to choose a proper intermediate layer in the backbone to incorporate the decoder. For better efficiency, you could save the visualization results of the predicted R during training. If your implementation works well, the visualized R would look good in the very early training stage.
Hope you find these helpful.