search

CU-Robosub / Yolo-Train-System

Convolutional Neural Networks
http://pjreddie.com/darknet/
0 stars 0 forks source link

Add Data Augmentation #11

Closed skhadem closed 5 years ago

skhadem commented 5 years ago

We don't want our model to overfit to the pool environment so we need some data augmentation, mostly of the brightness. Since it seems AlexyAB's training is a good way to train YOLO with darknet backend, the augmentation should be compatible with that method. The easiest way would be to generate some extra images per image with brightness varied and BB the same, and glob them all in the train directory. Ideally it should be randomized per batch but this would require modifying the code we use to train, which may not be a bad idea.

DrYerzinia commented 5 years ago

You know this is already built into darknet right? https://github.com/AlexeyAB/darknet/blob/55dcd1bcb8d83f27c9118a9a4684ad73190e2ca3/cfg/yolov3-tiny.cfg#L15

skhadem commented 5 years ago

Is that random or applied to every image?

skhadem commented 5 years ago

Oh yeah you are right

DrYerzinia commented 5 years ago

https://github.com/AlexeyAB/darknet/issues/644 For more details