PointNet goes PyTorch, but leaves some points on the table

What it does

Reimplements the PointNet architecture from the 2016 Stanford paper for 3D classification and segmentation directly on point sets. The model lives in pointnet/model.py, with training scripts for both tasks included and tested against PyTorch 1.0. It targets standard benchmarks: ModelNet40 for classification and a ShapeNet subset for part segmentation.

The interesting bit

The value is purely in framework translation: the original implementation was TensorFlow-based, and this offers a clean PyTorch alternative. The README deserves credit for transparency, openly showing that classification accuracy on ModelNet40 lags the original by a couple of points and that segmentation mIOU drops more dramatically on several ShapeNet classes.

Key highlights

• Direct PyTorch translation of the seminal PointNet paper
• Handles both 3D object classification and part segmentation
• Optional --feature_transform toggle for an architectural variant
• Includes a rendered sample segmentation result
• Segmentation classes are trained separately, which the README admits hurts performance on categories with less data

Caveats

• Accuracy trails the original implementation across the board; segmentation gaps are especially wide (e.g., Airplane mIOU of 73.5 vs. the original 83.4)
• The segmentation performance table leaves many feature-transform scores blank
• Only explicitly tested on PyTorch 1.0, so behavior on modern releases is an open question

Verdict

A sensible choice if you want a readable, self-contained PointNet reference in PyTorch and can absorb a modest accuracy penalty. If you need the exact original numbers or modern production performance, look elsewhere.