康行天下

You can try it, other thing like iou thresh.

The pre-trained model of backbone network such as ResNet and VGG is the same as corresponding paper. You can download those from the link written in README. The other trained...

since the crop layer has been removed,and if the width or height of image is not multiple of 2^6, the upsample layer will not get exactly same shape of output....

@unsky Yes, I didn't read the data layer codes before. you do resize by keeping the aspect ratio then use pad to align IMAGE_STRIDE. using padding is truely a good...

Hi, When I train the models such as `res50-12s-600-rfcn-cascade` without FPN with my own dataset is fine. But when I try to train `res50-15s-800-fpn-cascade` with my own dataset, I meet...

@zhaoweicai Thanks! Follow your advice, set lower the fg_thr in BoxGroupOutput layer, the problem disappeared.

@Peng-wei-Yu try lower the score of `fg_thr` instead of nms thresh.

@Peng-wei-Yu If you use the author's prototxt, you should use the corresponding ResNet-50-model-merge.caffemodel, since it merges the BN layer to scale layer to reduce memory and speed up. You can...

@GuoxingYan I set fg_thr: 0.01 or 0 in all BoxGroupOutput layer. If your positive rois num is always 0, maybe your dataset has some problem.

@GuoxingYan I did not try to change that, since there use Deconvolution layer to upsample, the size maybe need to be multiplier of 32, 64 or larger.