Extraction of flow based features
znbmalik opened this issue · 2 comments
Hi, I am trying to understand your code line by line but I did not work with lua torch so I am facing some difficulties to execute it. I am creating my model in tensor flow and seeking some pre-processing guidance from your model. Just like your work, I have also extracted flow-x and flow-y using Brox algorithm code (that your shared). The obtained values of flow-x and flow-y were in -ive to +ive range which have been shifted to 0-255. but I when I subtract mean values from each frame(as you mentioned in your work), I got -ive values again. My point of concern is that in CNN model ReLu is used as activation function that ignores -ive values which means that after all pre-processing my flow features should have only positive values but after subtracting mean values I got -ive values also for flow-x and flow-y which I stored as '.npy'. Can you please share which range of values you had in your final flow features before passing to CNN (I have no understanding of lua torch and thus don't know how to see the content of t7 files).
The input to the CNN is the raw flow value in terms of pixels (minus the mean flow per frame). For simplicity, the flow values were stored as 0-255 jpegs, but the min/max values that were used to scale the flow were also stored to be able to go back to the original flow data (see
https://github.com/gulvarol/ltc/blob/master/donkey.lua#L151-L169).
Thank You for your guidance. Really appreciated. If I correctly got your point then It means that before passing flows to CNN you are bringing the flow values to its original values that could have both +ive and -ive values.