I have also added a function to compute the total logprob of the image given the nnet output: compute_total_logprob(self)
I am not sure if I'm computing it right, e.g. whether we need to include object sameness probs for the pixels inside the object or not?
Additionally I have changed the formula for merge priority by dividing obj_merge_logprb by a constant c > 1.0. This leads to much better segmentations and better final total logprbs. I imagine c might be related to num-offsets. What would you think about this?
Hossein
I have also added a function to compute the total logprob of the image given the nnet output:
compute_total_logprob(self)I am not sure if I'm computing it right, e.g. whether we need to include object sameness probs for the pixels inside the object or not?Additionally I have changed the formula for merge priority by dividing obj_merge_logprb by a constant c > 1.0. This leads to much better segmentations and better final total logprbs. I imagine
cmight be related to num-offsets. What would you think about this? Hossein