search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.72k stars 7.96k forks source link

Too many false positive #2815

Open wsung1 opened 5 years ago

wsung1 commented 5 years ago

Hi, I'm trying to train a YOLO v3 model for the road traffic sign detection. I believe I've followed most of instructions kindly provided by @AlexeyAB but my results have too many false positives, as shown below. result1_image result1_terminal The training conditions include cfg I'd appreciate if you give any pieces of advice.

AlexeyAB commented 5 years ago

https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects change line batch to batch=64 change line subdivisions to subdivisions=8

wsung1 commented 5 years ago

As traces are left in the above .cfg file, I started my training with subdivisions=8 but I had no choice but to increase it to 64 due to out of memory error :(

AlexeyAB commented 5 years ago

Did you check your dataset by using Yolo_mark? How many iterations did you train? What avg-loss and mAP did you get at the end of training? Did you use cuDNN?

wsung1 commented 5 years ago

Yes, I checked. 40,000 iterations, resulting in an average loss of 0.1, as displayed below chart Sure, I trained with cuDNN 7.4.2.

ghost commented 5 years ago

@wsung1 i noticed that there is a difference in confidence depending on which version of darknet you are using ( pjreddie/darknet, alexeyab/darknet, C++ wrapper yolo_v2_class.cpp). It mays be that , chekc my issue :

https://github.com/AlexeyAB/darknet/issues/2481