Zhang Yifeng

Try to use the DepthwiseConvolution layer instead of Convolution type, if the group parameter is not 1. The DepthwiseConvolution implementation can be found [here](https://github.com/yonghenglh6/DepthwiseConvolution)

after resize the min side to 256 and center crop the image, the accuracy is 0.7189😀

i changed the linked library in Makefile, line 181, _hdf5_serial_hl hdf5_serial_ into _hdf5_hl hdf5_

so the batch size of 384x384 and 768x768 are both 32, is that right?

Try to change the strides of the 5th and 7th block from _s22_ back to _s11_, and extract the last layers of 1/2/3/5/7 block. Codes are [here](https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/master/models/utils.py#L267) and [here](https://github.com/toandaominh1997/EfficientDet.Pytorch/blob/master/models/efficientdet.py#L99) Remember...

@gmvidooly here's code ``` def extract_features(self, inputs): # Stem x = self._swish(self._bn0(self._conv_stem(inputs))) P = [] index = 0 num_repeat = 0 # Blocks for idx, block in enumerate(self._blocks): drop_connect_rate =...

if you not familar with cmake, just config the Makefile.config file by yourself. That maybe more controllable.

If you are using opencv3, please make sure this place should **not** be commented: https://github.com/unsky/FPN/blob/master/caffe-fpn/Makefile.config#L21

I use the released ckpt in this git, test on the ICDAR Incidental 2015 test set. The result seems not good enough. Recall/Precision/Hmean: 0.666/0.035/0.067 Using prob score, under the best...

The author says N in loss function is the number of default boxes that match groundtruth boxes, but in this code is the batch size instead.