This isn't really a theano question and might get more traction elsewhere.

Have you considered adding a "no action" class to your problem for the 
frames/times where nothing is happening?

On Thursday, September 15, 2016 at 8:15:01 AM UTC-7, aditya vora wrote:
> Hello , 
> We are training a 3D convolutional neural network with real life videos 
> that we recorded from cameras. Actually, my dataset has around 3000 video 
> clips which are obtained by dividing a single surveillance video. Each clip 
> was appropriately labelled by us. I need to classify this videos into 
> different actions. However, as the data is obtained by capturing real life 
> videos it is certainly not possible to have totally clean data, as a 
> result, there are certain situations in the video clips in which there are 
> no actions happening. However we tried to remove such clips as much as 
> possible but still couldn't filter out all clips because some frames of the 
> same clip had actions in it (Eg. Person running and going out of the 
> frame). So my question is how the convolutional neural network training 
> shall behave in such a situation? Will the network be immune to such 
> "outliers" as most of the data is appropriate? If it will affect the 
> training then how it shall affect and if not how? I need a mathematical 
> proof to this question in terms of parameter update and stuff.
> Thanks, 
> With Regards, 
> Aditya Vora    


You received this message because you are subscribed to the Google Groups 
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
For more options, visit

Reply via email to