cw

I pushed an issue just now, is the same problem as yours, I think it is no need to use bn here, cuz bn is somehow channel-wise normalization, thist require...

yep, if u only want to get the detected object, RPN is enough. But, in order to get the more precise results, it should combine with the detection head.

Here is an example to illustrate this problem: Consider the 1st point (0, 0) in feature map, due to the feat stride is 16, the center point of 16x16 region...

in train.py, u can freely define the batch size in train().

But I found if we set batch size > 1, it will raise an exception error..

it will take some works if u wanna train with batch size > 1. First, u should change the code for data processing, cuz each image has different number of...

The direct way is to comment this line: resource.setrlimit(resource.RLIMIT_NOFILE, (20480, rlimit[1]))

I got the same problem, setting os.environ['CUDA_VISIBLE_DEVICES'] doesn't work! have u solved it now?

算法内容也可以按岗位细分，比如CV、NLP、推荐甚至语音，有针对性地收集与整理一些内容。

IoU < 0.5 with all anchors is rarely seen, if this did happen, you may change the scale & aspect ratio setting of anchors.