Du Tran

At extract_image_features.cpp:76: it checks the network and your net does not have a layer with name "prob"

@mystao model parameters are just a small part of memory requirement. Most of the time memory was mostly consumed by activations (input/output blobs and column buffers). Thus, your input size...

It may be due to the cudnn supports, can you try the earlier one be for we merge the PR https://github.com/facebook/C3D/pull/327

Not sure I understand you correctly, but what I did is to extract feature for clips which 8-frame overlapped. Since a video has multiple clips, we average pool these clips...

I think Makefile needs to be revised to support CUDA 9.0.

Sorry, I did not have time to do that, since Scene datasets are all small, I guess you can do that and, if not, someone else can help?

@sherleens Can you check if you turn of batch-norm to to use_global_statistics: true, v1.0 has no batch norm while 1.1 does

It looks like you got linking issue, some of them from protobuf. which may cause by verion miss-match.

your video filename has some special characters that may cause the issue.

@ltnghia I just ported deconv3d layer. I'm currently busy with a deadline, I will update the data layer later.