search

HuguesTHOMAS / KPConv-PyTorch

Kernel Point Convolution implemented in PyTorch
MIT License
786 stars 157 forks source link

Why is in_features_dim = 5 for S3DIS? #245

Closed DarylWM closed 6 months ago

DarylWM commented 7 months ago

Thank you for sharing this project; I'm finding the code to be clear and well commented, which makes it easier to understand.

One parameter I don't undestand though is why the input features dimension is 5 for the S3DIS dataset. Why would it not be 6, for (x,y,z,r,g,b)?

Daryl.

LucaRom commented 6 months ago

Hi, the features_dimension value does not include the x, y, z coordinates, it's for the number of extra features to include in the training. For example S3DIS as R,G,B,D (depth layer) and KPConv adds a feature of 1's to all training by default.

Related answer by the author : https://github.com/HuguesTHOMAS/KPConv-PyTorch/issues/156#issuecomment-1081892752

DarylWM commented 6 months ago

Thank you very much. I had missed that previous answer.

HuguesTHOMAS commented 6 months ago

Just to correct a small mistake, the 4th is not depth but the z coordinates (to encode height above ground). In a setup where the inout is randomly augmented with horizontal rotations, x and y are not very informative, but z remain a meaningful feature