search

isl-org / Open3D-ML

An extension of Open3D to address 3D Machine Learning tasks
Other
1.88k stars 321 forks source link

Documentation #376

Closed kukuruza closed 3 years ago

kukuruza commented 3 years ago

Is there a documentation on the architecture somewhere? I am trying to study the codebase, but there are few comments, and there are no docs.

In particular, I'm trying to understand one simple thing - when training / testing the semantic segmentation pipeline (with randlanet or kpconv models) on SemanticKITTI, do point clouds from several samples in the same sequence get merged together, or are they completely independent?

kukuruza commented 3 years ago

In particular, I'm trying to understand one simple thing - when training / testing the semantic segmentation pipeline (with randlanet or kpconv models) on SemanticKITTI, do point clouds from several samples in the same sequence get merged together, or are they completely independent?

Seems like the answer is INDEPENDENT for RANDLANET because the file poses.txt is not used, and it is impossible to merge point clouds without this file.

sanskar107 commented 3 years ago

For RandLANet, we don't merge different pointclouds. While KPConv doesn't have any batch dimension, and different pointclouds in the same batch are concatenated. The nearest neighbors are computed keeping this in consideration.