Directly Using FlowNet2_weights.caffemodel.h5 without finetuning.But optical flow is blank.

[gsygsy96]

I doubt that I didn't config caffe correctly. 1.make all -j8 2.make pycaffe Lacking 'make tools -j8' will make difference with testing? By the way, when I run'make tools ', " make: Nothing to be done for 'tools' "shows out.Although bin file doesn't exist.

[gsygsy96]

I have tried figures in data/FlyingChairs_examples.Optical flow is still blank.

[nikolausmayer]

make all includes tools, so you don't actually have to build it separately. Our instructions are a bit, uhm, in need of optimization there... ;)

I think we need a bit more information. What commands did you run? What was the output? Which files were created? How do you display the flow files?

[gsygsy96]

commond: python scripts/run-flownet.py models/flownet2-models/FlowNet2/FlowNet2_weights.caffemodel.h5 models/flownet2-models/FlowNet2/FlowNet2_deploy.prototxt.template 0000000-img0.ppm 0000000-img1.ppm 111.flo

output:111.flo

diaplay flow files: visualize_flo.py.zip

I can show grountruth in'flownet2/data/FlyingChairs_examples'.So I think code is right.

[nikolausmayer]

I'm getting this output, so the network produces a result (I'm probably using a wrong color scheme). Does the network print an error message? Try the --verbose option!

[gsygsy96]

No error message.How to config --verbose?

2017-11-30 16:25 GMT+08:00 Nikolaus Mayer notifications@github.com:

I'm getting this output, so the network produces a result (I'm probably using a wrong color scheme). Does the network print an error message? Try the --verbose option!

[image: bildschirmfoto_2017-11-30_09-22-33] https://user-images.githubusercontent.com/7045923/33420777-25212fca-d5b0-11e7-84f0-4d11ac6211f1.png

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/lmb-freiburg/flownet2/issues/93#issuecomment-348116423, or mute the thread https://github.com/notifications/unsubscribe-auth/AV2vwqVaXnokRB_Mol7k9ggoOf-7Ez1Jks5s7maMgaJpZM4QuoV3 .

[nikolausmayer]

python scripts/run-flownet.py --verbose models/flownet2-models/FlowNet2/FlowNet2_weights.caffemodel.h5 models/flownet2-models/FlowNet2/FlowNet2_deploy.prototxt.template 0000000-img0.ppm 0000000-img1.ppm 111.flo

[gsygsy96]

Ummm....Still not work.Maybe it due to enviorment ?My friend have run it correctly in server. And I have another question about two stream CNN.How can I stack 5 optical flow to one optical flow?Just average each dx and dy?

2017-11-30 17:02 GMT+08:00 Nikolaus Mayer notifications@github.com:

python scripts/run-flownet.py --verbose models/flownet2-models/ FlowNet2/FlowNet2_weights.caffemodel.h5 models/flownet2-models/ FlowNet2/FlowNet2_deploy.prototxt.template 0000000-img0.ppm 0000000-img1.ppm 111.flo

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/lmb-freiburg/flownet2/issues/93#issuecomment-348124691, or mute the thread https://github.com/notifications/unsubscribe-auth/AV2vwuBShpfCatHeG39wxTc-TBNV0zV2ks5s7m8QgaJpZM4QuoV3 .

[nikolausmayer]

That sounds like a good starting point to investigate. Maybe check version differences of CUDA, CuDNN etc between your machine and that server? :)

I'm not sure what you mean. Wherever we have multiple network inputs, they are all separate input channels (so 10 channels for 5 flow fields, for example). Averaging flow fields channel-wisely is probably not a good idea as it breaks the correlation between x and y...

[nikolausmayer]

(closed due to inactivity)

[zhanghaozhou]

@mehameha998 hi,

Did you solve this problem of blank flow? I got the same error. Thanks a lot for your help!!

Best regards, Haozhou