Advice for hyperparameter search?

[rgov]

The code in this repo and the thorough instructions are great. I'd like to see if I can increase the quality of detection on my data set. I was thinking about searching across the hyper parameter space for a parameter set that works a little better.

Are there particular hyper parameters I should focus on? For instance, the image resizing?

In my case, my data has a number of adverse properties (low light, motion blur, etc.) To augment the data first I run it through algorithms to segment the foreground, and I dedicate one channel to optical flow data. So the images look different than most.

For example (ignore the bounding boxes which are not correct due to an earlier bug):

[AlexeyAB]

I dedicate one channel to optical flow data.

• I think you should disable color data augmentation, comment lines with hue, exposure, saturation in cfg file and train

• How many channgels in your image?

[rgov]

There are 3 channels;

• Red - original grayscale image
• Green - foreground extraction bitmap (255 = foreground, 0 = background)
• Blue - optical flow magnitude

[AlexeyAB]

comment these lines: https://github.com/AlexeyAB/darknet/blob/342a8d1561c19317f2d5fda0f099449b79b51716/cfg/yolov3.cfg#L14-L16

[rgov]

Thanks! Do you recommend searching any other parameters, such as width and height?

[AlexeyAB]

If all your trainin and test images have the same size 400x960, then use [net] width=416 height=960