[Help Wanted]Problem while training on customize datasaet

[laplaceson]

Hi @zylo117 , First thanks for your awesome work! I try to use pretrained weight D0 model on my custom dataset. However, it always gives reg loss 0 . So i try to use other method to modify the anchors_scales and anchors_ratios. (Since my object are small and sparse). The reg loss goes non-zero and converge. But i figure out my validation reg loss and training reg loss have big difference. And finally i use debug mode (Or using weight to efficientdet_test.py ) to visualize it. But get nothing detected. Do you have any suggestion? Thanks By the way, my task is single class object detection(object size often not big).

[laplaceson]

Another question, i see some of your comment says it's hard to detect smaller 32 32 , is this means that under 512 512 's image, if the annotation is smaller than 32 * 32, it's hard to detect in this network? Thanks

[zylo117]

32x32 under 512x512