0

I have trained a custom YoloV4, single class object detector using AlexeyAB's darknet.

The best weight files of my model result in the following performance

metrics of my model trained on google colab

I obtained this evaluation using the command,

darknet.exe detector map data/obj.data yolo-obj.cfg backup\yolo-custom-best.weights

Upon running the model on a video, I get around 25 fps.

My Question is : When I read articles about YOLO, they claim that it detects in milliseconds (ms). Why is my model showing detection time of 5 seconds but comparable in FPS to other, much faster models?

  • Is the detection time shown in the image (returned by the given command), something different? (If yes, then how do I find how fast my model is?)
  • If No, how to make it faster?
varungupta
  • 155
  • 4
  • 10

1 Answers1

1

Please note that the detection time mentioned in the paper, which is the same as the one mentioned in the GitHub repo here is assuming it's run on a single batch at test time on a Tesla V100 GPU. In addition to these factors, the runtime also varies according to the size of the input image

In case you are using a different GPU or the CPU, it shouldn't be surprising for you to find longer detection times. The same goes with the FPS

Jitesh Malipeddi
  • 2,150
  • 3
  • 17
  • 37
  • So I ran a detection with my model, and there too the time gets mentioned with detection, it's around 30ms. So, I guess this time is of something different, though it says detection time. – varungupta Apr 21 '21 at 16:01