Foreword
Previous I have introduced the use of pre-trained yolov3 weights file (containing only convolution layer) and training
All you need to compile:
./darknet detector train cfg/voc.data cfg/yolov3.cfg darknet53.conv.74
While the weights will generate a corresponding backup file folder (file update with increasing time constant), as follows:
1. Modify the parameter file (../darknet/cfg/yolov3.cfg)
2. Re-make under the darknet
make # recompile
3. Run the following command
./darknet detector test cfg/voc.data cfg/yolov3.cfg backup/yolov3_300.weights -thresh 0.1
Description: a weighting name modification; 2.YOLOv3 default threshold is 0.25, -thresh 0 is set thresholds set to 0, then you can show all test results; 3 Operation End Direct will let you enter. Enter Image Path:
4. The significance of training process parameters
Region xx: cfg file yolo-layer index; Avg. The IOU: the current iteration, the predicted average deposit box marked box and the ratio, the better, as the expected value. 1; Class: label object classification accuracy, the bigger the better, as the expected value. 1; obj: the bigger the better, as the expected value. 1; No obj: the smaller the better; .5R: IOU = 0.5 to the threshold value when a recall; recall = positive samples detected / actual the positive samples 0.75R: IOU = 0.75 in the threshold value when the Recall; COUNT: n number of samples.
Reference blog: