FPS is too slow?

[jinertianzhenhao]

I have test the CornerNet with one 1080ti. The fps of CornerNet-Squeeze is 16.4 while the fps of yolov3 is 26.3,so the speed is too slow.
Is there any mistake in the steps I made my prediction? The command I tested was:
1.python evaluate.py CornerNet_Squeeze --testiter 500000 --split validation
2.Then divide 1000ms by the average time displayed to get fps
How do you test fps?

[heilaw]

On our machine with a 1080Ti and Core-i7 7700k, the average FPS of CornerNet-Squeeze is 33fps while that of YOLOv3 is 30fps. The difference in the inference time can be due to the difference in software setup. Are you using Python 3.7 from Anaconda? Which version of CUDA are you using?

[jinertianzhenhao]

I used a 1080ti with Python 3.7 and cuda10. Is speed related to the frequency of cpu or something else?
And the code that gets the average time is the same as in the project, I made no changes.
timer.tic()
top_bboxes[image_id] = cornernet_inference(db, nnet, image)
timer.toc()

[jinertianzhenhao]

In addition, I found that you are describing the average fps. What is the average of what you are taking?

[heilaw]

The speed can also depend on the frequency of CPU since part of the code is executed on CPU. However, I am not sure how much it would affect the inference time.

The average FPS means that I divide 1000ms by the average time displayed, which is the same as your step 2.