search

jzbontar / mc-cnn

Stereo Matching by Training a Convolutional Neural Network to Compare Image Patches
BSD 2-Clause "Simplified" License
707 stars 232 forks source link

Apply trained network to image of other resolutions #49

Open wzrf opened 6 years ago

wzrf commented 6 years ago

Hello! I want to know how can I apply the trained network to images with other resolutions? I notice it's 1266x370, how can I apply this to middlebury dataset?

wzrf commented 6 years ago

I also find that the example input are gray-scale images, I want to know what kind of preprocess I should do on colorful images before I use them on the network?

LUCASLLA commented 6 years ago

hi, have you discovered anything? I tried to change the images on the samples folder to some of my images, but with the same format and resolution (1266x370) and also in grayscale, and I`m having the error:

luajit: ./main.lua:1124: bad argument #2 to '?' (sizes do not match at /home/lucas/torch/extra/cutorch/lib/THC/generic/THCTensorCopy.c:48) stack traceback: [C]: at 0x7f7d610de490 ./main.lua:1124: in main chunk [C]: at 0x00405d50

Tks.

saeid-h commented 6 years ago

That won't be a problem if your left and right images are in the same size and "png grayscale". I've tried for different image sets. Use the given command line and replace your image path and correct disparity: ./main.lua kitti fast -a predict -net_fname net/net_kitti_fast_-a_train_all.t7 -left left.png -right right.png -disp_max disp