PointNet++ trains harder, hits state-of-the-art

What it does

PointNeXt is the official PyTorch implementation of a NeurIPS 2022 paper that revisits PointNet++. It ships inside OpenPoints, a modular research library for point cloud understanding, and demonstrates that much of the performance gap between classic point-based networks and modern leaders comes down to training and scaling strategies rather than architecture alone.

The interesting bit

The authors trained every implemented model at least three times and report mean±std, a level of rigor that remains depressingly uncommon. They also retrained existing architectures—PointNet, DGCNN, DeepGCN, and others—with their improved recipe and found they consistently outperformed their own originally published scores, suggesting the 3D vision leaderboards have been partly benchmarking who bothered to tune hyperparameters.

Key highlights

• OpenPoints supports a wide roster of point cloud networks: PointNet, DGCNN, DeepGCN, PointNet++, ASSANet, PointMLP, and PointNeXt
• All models are trained at least three times; the authors publish mean±std, pretrained weights, and logs
• Claims fair benchmarking by applying the same improved training strategies across all models, lifting accuracy above original reported values
• Configuration-driven: models, optimizers, schedulers, and loss functions are assembled from YAML files
• Integrates wandb for experiment tracking and maintains a model zoo for classification and segmentation tasks

Verdict

Grab this if you are doing point cloud research and want a reproducible, benchmark-conscious training framework with multiple baseline architectures ready to go. Skip it if you need a polished production runtime or a consumer application.