search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

Flow Net with MatConvNet #337

Closed Monicar77 closed 8 years ago

Monicar77 commented 8 years ago

Hello!

I am trying to implement FlowNet network from this paper FlowNet: Learning Optical Flow with Convolutional Networks, ICCV 2015.

The authors propose to use two different modules in the network. The first one is a usual CNN. The second one is a deconvolutional network that provides the optical flow from the input images as output.

Regarding deconvolutional networks I am quite confused on how to implement it. Would it be sufficient with MatConvNet convt layer or should I implement also the uppooling and the inverse relu operations?

Thank you very much!

vedaldi commented 8 years ago

Hi,

I am trying to implement FlowNet network from this paper FlowNet: Learning Optical Flow with Convolutional Networks, ICCV 2015.

The authors propose to use two different modules in the network. The first one is a usual CNN. The second one is a deconvolutional network that provides the optical flow from the input images as output.

Regarding deconvolutional networks I am quite confused on how to implement it. Would it be sufficient with MatConvNet convt layer or should I implement also the uppooling and the inverse relu operations?

I think convt is all you need. As far as I know, what they mean by “inverse relu” is just implemented as a relu (but I could be wrong).

Andrea

Thank you very much!

— Reply to this email directly or view it on GitHub https://github.com/vlfeat/matconvnet/issues/337.

alijavidani commented 6 years ago

@Monicar77 : Hi, I was wondering did you succeed in implementing FlowNet with Matconvnet?

If yes, would you please share the code?

Best Regards,

Ali

Monicar77 commented 6 years ago

Dear Andrea,

As far as I can remember, the implementation of FlowNet required to perform deep modifications on some layers. Therefore I finally installed Cafe on my computer and used the codes provided by the authors. I tested the method on some sequences of my interest and the results were pretty bad compared with the state of the art methods. Of course, I believe that the results wold have improved with an specific training but I did not have enough time to perform such work.

You can have a look to the results in (the results are gathered on a table)

https://www.researchgate.net/publication/321273945_Primal-dual_optimization_strategies_in_Huber-_L_1_optical_flow_with_temporal_subspace_constraints_for_non-rigid_sequence_registration

I am sorry for not being more helpful but deep learning is not my main research focus (yet).

Best regards,

Monica.

2018-03-18 14:15 GMT+01:00 alijavidani notifications@github.com:

@Monicar77 https://github.com/monicar77 : Hi, I was wondering did you succeed in implementing FlowNet with Matconvnet?

If yes, would you please share the code?

Best Regards,

Ali

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/vlfeat/matconvnet/issues/337#issuecomment-373998083, or mute the thread https://github.com/notifications/unsubscribe-auth/AOhkAlh_a9pF2ZgXSvnJKaL0lknMELhdks5tfl37gaJpZM4GuF5P .

-- Monica Hernandez Gimenez Associate Professor Computer Sciences Department University of Zaragoza, Spain