yolov2 on 1 class difficulty

[aniketvartak]

Hi @AlexeyAB , I am training yolov2 on 1 class. I have about 10k images annotated. the training outputs following log:

Tensor Cores are used. 53: 0.645953, 2.186047 avg loss, 0.000100 rate, 0.099066 seconds, 212 images MJPEG-stream sent. Loaded: 0.065090 seconds Region Avg IOU: 0.546212, Class: 1.000000, Obj: 0.055733, No Obj: 0.055440, Avg Recall: 1.000000, count: 1 Region Avg IOU: 0.578892, Class: 1.000000, Obj: 0.055941, No Obj: 0.055564, Avg Recall: 1.000000, count: 2 Region Avg IOU: 0.533764, Class: 1.000000, Obj: 0.058952, No Obj: 0.055555, Avg Recall: 1.000000, count: 1 Region Avg IOU: 0.428985, Class: 1.000000, Obj: 0.056616, No Obj: 0.055440, Avg Recall: 0.000000, count: 2

where the Obj and No Obj after some time goes to 0.0000. like: Loaded: 1.011449 seconds Region Avg IOU: 0.599942, Class: 1.000000, Obj: 0.000000, No Obj: 0.000000, Avg Recall: 0.736842, count: 19 Region Avg IOU: 0.580819, Class: 1.000000, Obj: 0.000000, No Obj: 0.000000, Avg Recall: 0.739130, count: 23 Region Avg IOU: 0.549266, Class: 1.000000, Obj: 0.000000, No Obj: 0.000000, Avg Recall: 0.625000, count: 24 Region Avg IOU: 0.554060, Class: 1.000000, Obj: 0.000000, No Obj: 0.000000, Avg Recall: 0.695652, count: 23 I train this for few thousand epochs and the loss goes down to 0.xxx

when I test the saved model nothing is detected. I have sampled my image annotations to check and the annotations are ok.

I want to understand what does Class: 1.000000, Obj: 0.000000, No Obj: 0.000000 means and how I can improve?

[AlexeyAB]

@aniketvartak Hi,

• What mAP can you get?
• Try to run training with flag -show_imgs do you see correct labels?

I want to understand what does Class: 1.000000, Obj: 0.000000, No Obj: 0.000000 means and how I can improve?

https://github.com/AlexeyAB/darknet/issues/636#issuecomment-381400954

[aniketvartak]

unfortunately, I am using GCP so can't do -show_imgs mAP is 0.

any idea?!

[AlexeyAB]

Check your dataset by using Yolo_mark: https://github.com/AlexeyAB/Yolo_mark