search

GOATmessi8 / RFBNet

Receptive Field Block Net for Accurate and Fast Object Detection, ECCV 2018
MIT License
1.41k stars 355 forks source link

Speed of RFBNet and SSD on COCO and VOC, have you consider the nms time? #92

Closed tjulyz closed 5 years ago

tjulyz commented 5 years ago

Hi, Thank you for share your codes! When I test your code on VOC and SSD, I cannot reproduce the speed as you reported. My configuration is Titan XP, pytorch 0.4, cuda 9.0, and get 20+ms for VOC, 30+ms for COCO, not like yours (80+fps, and 15ms for COCO). When test the speed, I use conf_thresh 0.01, gpunms. I have also tried pytorch 0.3, still cannot reproduce the speed. The time of nms for me is ~20ms for COCO, and ~10ms for VOC. So have you consider the NMS time when report the fps? Any advice for getting the correct speed? Thank you very much!

GOATmessi8 commented 5 years ago

@tjulyz In my test, I get ~9ms inference speed plus ~3 ms CPU nms time with pytorch 0.3.1, cuda 9.0 on VOC 300. For COCO 300 the GPU nms time increase to ~6ms. These measurements all ignore the first 50 samples when the GPU is warming up. For your configuration, I guess the bottleneck mainly comes from the nms. You can try the gpu nms version or use a higher version of numpy and cython, or a more powerful cpu.

tjulyz commented 5 years ago

Great, thank you very much!

vaesl commented 5 years ago

I also met the same problem. My configuration is Titan XP, pytorch 0.3.1, cuda 9.0 and I am using the nms of gpu version. For COCO 300, the inference speed is ~6 ms/img while the nms time is ~12ms/img, which looks strange. Could you please provide some more suggestions? Many thanks!

GGjerry01 commented 5 years ago

i alse met the same problem, gpu_nms also need cost much time than model infer. i am very worried about this