search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
959 stars 206 forks source link

Augmentations #40

Closed padus94 closed 3 years ago

padus94 commented 4 years ago

Thank you for this great work! It's an amazing platform to understand different CNN-architectures for 3D-CNN and segmentation.

I try to understand the effects of augmentations on the projections and observe the IoU's, but for myself it's hard to understand where it is best to implement the effects/noise on the trainsequences. It was possible for me to make some changes on the perspectives and image size through the Spherical Projection class, but these effects are also changing the validation set.

In which programmpart (user/parser/ect.) can i find the step where the training tensors are put inside the decoder?

Thank you !

tano297 commented 4 years ago

Hi there, Thanks for the props. If I understand what you want to do correctly, the best place is in the function __getitem__ here. This is using the pytorch Dataset class, so getitem is called by the generator that gives you batches in the training loop. The index is a random index from 0 to Nscans. Each call to getitem is supposed to return one training example, and they are all batched by the DataLoader included in Pytorch. Does this answer the question?

padus94 commented 4 years ago

Yes! Thank you for your prompt reply!

jbehley commented 3 years ago

I close this issue since there seems not to be much activity here or the problem resolved.