Open pkdogcom opened 7 years ago
GOTURN tracker uses dnn module to perform forward propagation. There are no plans to use CUDA as a backend in DNN, but some other performance improvements will be coming soon. It's better to ask @Auron-X about tracking quality issues, but I think all you can do to get better results is to train GOTURN network on your data.
any updates on how to utilize GPU for goturn rather than cpu ?
Hi! DNN module now supports Intel GPUs and NVIDIA GPUs. So now it's a question of API of how to switch target for dnn network.
The GOTURN tracker doesn't seem to be utilizing CUDA to perform forward propagation, and thus the performance is really bad. Also, bounding box of the current implementation seems to keep enlarging without correctly track on target. The prototxt and caffemodel that I used are from opencv_extra repository: https://github.com/Auron-X/opencv_extra/blob/3d5551b4468f212fd727d32c0bb7bda9302168ce/testdata/tracking/goturn.prototxt