Skip to main content
Back

LIDARLearn Unifies 55 Point Cloud Models Where Every Other Library Failed

A new open-source PyTorch library unifies 29 supervised backbones, 7 self-supervised methods, and 5 fine-tuning strategies for 3D point cloud research.

Pro

  • LIDARLearn unifies 29 supervised, 7 SSL, and 5 PEFT point cloud methods in one PyTorch library.
  • Registry-based YAML configs drive shared runners for classification, segmentation, part-seg, and few-shot tasks.
  • Ships stratified K-fold CV, Friedman/Nemenyi tests, and 2,200+ automated tests validating every configuration.
  • SSL implementations were dual-audited against original repositories for loss, checkpoint, and freezing behavior.
  • Case study on STPCTLS tree species: ACT + PointGST reaches 69.78% accuracy training just 0.62M parameters.
  • Released under the MIT license; paper and code are freely available.

Anyone who has tried to benchmark two point cloud models from different papers knows the pain: mismatched data loaders, conflicting CUDA extensions, evaluation scripts that report subtly different metrics, and YAML formats that share nothing but the file extension. A new library called LIDARLearn tries to end that fragmentation by putting the modern 3D point cloud stack behind a single registry-based interface.

One repo to rule the point cloud zoo

LIDARLearn is a PyTorch library that integrates over 55 model configurations covering 29 supervised architectures, seven SSL pre-training methods, and five PEFT strategies, all within a single registry-based framework supporting classification, semantic segmentation, part segmentation, and few-shot learning. Prior open-source libraries only cover slices of that surface. The authors position their release as the broadest-coverage unified PyTorch library for 3D point cloud deep learning.

Self-supervised pre-training (SSL) here means learning point cloud representations from unlabeled 3D shapes, typically by masking random patches and reconstructing them, as popularized by Point-MAE and Point-BERT. Parameter-efficient fine-tuning (PEFT) means adapting those pre-trained backbones to a new task by training only a small set of adapters or prompts while the base weights stay frozen, borrowing the idea from LoRA-style tuning in NLP.

Why existing toolkits fall short

The paper compares directly against five established libraries. The gaps around SSL and PEFT are stark:

LibraryClsSegPartSSLPEFT
Learning3D