Started working on an improved data set for YOLO training. I dropped the image size from 608x608 to 416x416, and instead of doing a single crop from each BDD training image, I wrote a crawler that does 54 separate crops. As you can see, this moves the crop around the original image and repositions the labelled objects, resulting in at least an order of magnitude increase in unique grid cell locations getting populated with data (even though I also reduced the grid to 13x13, still have 10x more cells populated).
.
.
.
