I have used detectron2 with RCNN R50-FPN model for one-class segmentation task. I have a problem with slow inference - I get 0.2s/image on gpu, but the documentation says 0.043s/image or even less is possible. My images are big (about 8000x2000), but there is resize in predict function.
I tried to add some power to my gpu, but results are the same. I also changed predict function to work on image as tensor instead of numpy and then pytorch conversion and it has helped, but I still can't get results like from documentation. Could you give me some advice on that? Maybe there are some parameters in config I should change? Now I have default settings.