search

szagoruyko / cvpr15deepcompare

Code and models for "Learning to Compare Image Patches via Convolutional Neural Networks"
http://imagine.enpc.fr/~zagoruys/deepcompare.html
Other
468 stars 169 forks source link

caffe model output #5

Open edgarriba opened 8 years ago

edgarriba commented 8 years ago

is it correct that the number of outputs is 1 for all the caffe models?

szagoruyko commented 8 years ago

I did not convert descriptor models assuming that its easy to edit prototxt file and get it

edgarriba commented 8 years ago

hey! thanks for your answer but I'm not sure if I understood it. Does it mean that you haven't test it? After running Torch models the resultant descriptors have size 512. Last question about the benchmark, did you proper normalize patches regions before fetching the networks using VL_COVDET() with 'Patch' flag? URL Thanks in advance!

szagoruyko commented 8 years ago

@edgarriba caffe networks are tested against torch on conversion. Regarding VL_COVDET I don't remember, have to dig in the code. will answer later.

edgarriba commented 8 years ago

nice! I'm struggling a bit to get proper results with your network. In the following plot I've tried some feature extractors and injected the results to vlfeat benchmark in the grafitti dataset. graf In this case MSER siam and siam2stream results come from the *.bin files you provide along with the raw patches obtained with MSER+VL_COVDET. MSER Alexnet is a caffe model in caffe repo. MSER iri is the model from this paper http://cvlabwww.epfl.ch/~trulls/pdf/iccv-2015-deepdesc.pdf

Thanks again!