Training outputs in vrcnet

paul007pl / VRCNet

[CVPR 2021 Oral] Variational Relational Point Completion Network

https://paul007pl.github.io/projects/VRCNet.html

MIT License

156 stars 29 forks source link

Training outputs in vrcnet #20

Closed stefano-mazz closed 2 years ago

stefano-mazz commented 2 years ago

Hi, i have a question about the output dimensions during training. Is it true that fine and course ptc have dimension [2*batch_size, numberOfPoints, 3], since the decoder during training has twice the features and twice the partial cloud as input? I'm trying to implement it on another dataset and i do not understand it. Thank you in advance.

paul007pl commented 2 years ago

Hi, you are right. During training, we try two parallel pipelines: 1. one for reconstruction; 2. one for completion, and that is the reason you found the batch_size is double.

stefano-mazz commented 2 years ago

So, it is used only for the loss computation. Ok i understand thank you!