Open shuxiao9058 opened 6 years ago
@shuxiao9058 , You can try force_color: true For example : https://github.com/eric612/MobileNet-YOLO/blob/master/models/yolov3_coco/mobilenet_yolov3_lite_train.prototxt
but the most dataset is grayscale,and the rgb channels are equal,the training result not well now.
I think convert all images to gray,and train with gray is best.
If the most images are grayscale , you can try to turn off image distortion , like hue,saturation , these data augmentation may confuse training
and also train with force color?
can i train with sigle channel(grayscale image),and maybe the training is faster.
Yes , I think it can get best performance between speed and accuracy as the few color images not much help training
yes,i will try.thanks for your reply.
another question, how to modify the model you provide to use grayscale for training?
does caffe has the force_gray param?
Sorry , I have not tried this way , you can try search same requirement as ssd caffe.
How to work with the mobile-yolo model?convert all the image to grayscale and then process with the model?