Xiaolong Wang

I think it will be very easy to apply it on the faster rcnn on this repo by just changing the prototxt, but i don't have time to try it...

Hi guys, sorry for the late reply. The codes is pretty robust itself, you should get the same result if you just run the script.

Currently we have only released the code for ASDN. Note that we originally implemented ASDN and ASTN in torch for the paper, where the STN is much easier to implement....

please follow the installation instructions in py-faster-rcnn code (https://github.com/rbgirshick/py-faster-rcnn), which is: install cython and cd $FRCN_ROOT/lib make

Just see it as an easy way to increase training samples.

@icodingc that is because we are using cosine distance, which means after normalization layer, ||x|| = 1, ||x1|| = 1. Thus loss = max(0, (2-2x_x1) - (2-2x_x2) + margin )

Sorry, I did not see this earlier, it is like this: E5K-WUqpmvw/pairs2/0000001_1.jpg 1 E5K-WUqpmvw/pairs2/0000001_2.jpg 1 E5K-WUqpmvw/pairs2/0000002_1.jpg 1 E5K-WUqpmvw/pairs2/0000002_2.jpg 1 E5K-WUqpmvw/pairs2/0000003_1.jpg 1 E5K-WUqpmvw/pairs2/0000003_2.jpg 1 ... DmeLCEonmfk/pairs2/0000002_1.jpg 82902 DmeLCEonmfk/pairs2/0000002_2.jpg 82902 sj_uxLHFGMQ/pairs2/0000001_1.jpg 82903...

@bbrattoli can you try "tar xzf" with "collect.tar" and "tar xf" for "collect2.tar.gz"? the file suffix might be wrong.

convert_normal_test_loc.cpp, cropDet.cpp is in the folder "examples/seg" test_net_loc.cpp is in the folder "tools" They should be compiled using the Makefile