Planar parameters are not actually calculated. about indoor-sfmlearner HOT 5 CLOSED

@sconlyshootery
In our experiments, the planar parameters are not zeros. It may be zeros at the beginning of the training if the predicted depth is zeros. I think you can use the pretrained model the check this.

x_T_dot_x is used. It was added to x_T_dot_X according to super pixel segmentation. Please not x_T_dot_x and x_T_dot_X are different. Also, point_channel is the channel of sum_points where sum_points refer to P_n^T Y_m in the paper. Then x_T_dot_X (P_n^T dot P_n in the paper) and sum_points was combined to calculate the plane parameters.

from indoor-sfmlearner.

@sconlyshootery

scatter_add_ is an inplace operator. It will write the value to sum_points. If it's not the case in your environment, you may update PyTorch.

from indoor-sfmlearner.

I see, thank you for your kind explanation. Another question, is it normal to get nan in planar depth, it is related to bad prediction of depth?

from indoor-sfmlearner.

I see, thank you for your kind explanation. Another question, is it normal to get nan in planar depth, it is related to bad prediction of depth?

The nan values are not from your codes, I use planar loss for my purpose, I don't use patchmatch, whether lack of it will lead to this phenomenon.

from indoor-sfmlearner.

I think planar loss needs good depth initialization. You can add this loss when the predicted depth is reasonable.

from indoor-sfmlearner.