This repository contains the source code for the semantic image segmentation method described in the ICCV 2015 paper: Conditional Random Fields as Recurrent Neural Networks. http://crfasrnn.torr.vision/
We initialize the layer MultiStageMeanfield by filling the diagonal with specified values (e.g., 3 and 5 for spatial and bilateral filters, respectively and -1 for compatilibity transform matrix). And ideally, only these weights (i.e. the values on diagonal) will be updated during the training. However, I just find the values off diagonal are also somehow modified (e.g. 0.0 -> 0.1).
I was wondering if this is reasonable. What does it mean if so? Should we clean the off-diagonal values before each forward pass?
Dear @bittnt ,
We initialize the layer
MultiStageMeanfield
by filling the diagonal with specified values (e.g., 3 and 5 for spatial and bilateral filters, respectively and -1 for compatilibity transform matrix). And ideally, only these weights (i.e. the values on diagonal) will be updated during the training. However, I just find the values off diagonal are also somehow modified (e.g. 0.0 -> 0.1).I was wondering if this is reasonable. What does it mean if so? Should we clean the off-diagonal values before each forward pass?
Sorry I might be missing something.
Thanks! Kun