Superpoint Transformer - DALES Point Cloud Segmentation

Exact checkpoint

Selected artifact: spt-2_dales.ckpt in the authors' Zenodo record 8042712, version 1. The record lists approximately 2.9 MB and MD5 c658329e6d1a60a746cef8c62ca94a1d. This is the semantic SPT model, distinct from SPT-nano, SuperCluster and EZ-SP.

How it works

Preprocessing groups nearby points into a hierarchy of superpoints. Attention uses relationships between these groups at multiple scales; semantic predictions can then be transferred to points.

Typical input and output

Use the DALES dataset adapter and matching preprocessing configuration. The output represents semantic classes for an aerial LiDAR scene. Retain the partition-to-point mapping and original survey coordinates for full-resolution export.

Strengths

A relevant comparison for aerial mapping because an official DALES checkpoint is available. Group-level processing provides a different approach from learning directly over many individual neighborhoods.

GIS and remote-sensing use

Use the DALES configuration as a reference for separating scene classes in aerial surveys. Before mapping a new location, check label definitions, density, input attributes and acquisition differences against your target cloud.

Limitations

Partition quality affects boundaries and small objects. Superpoints are processing groups, not guaranteed building or tree instances. The repository has breaking changes and checkpoint-migration guidance: match a compatible release or follow its documented conversion path before loading this 2023 artifact.

Training and fine-tuning

The documented semantic/dales experiment supports training and evaluation with a selected checkpoint. Adapt the data module for custom clouds and labels; inspect partition quality before fine-tuning. Validate on separate survey areas and preserve preprocessing, class mapping and code revision alongside the resulting checkpoint.

Framework and hardware

PyTorch, Lightning and compiled point-cloud dependencies in the official Linux setup. The repository documents multiple GPU configurations; choose a recipe suitable for available memory. Review logging settings before training, since the default can send logs to an external service.

Files, license and download

Use the official Zenodo record to obtain spt-2_dales.ckpt and verify its published checksum. The source code is MIT licensed; the checkpoint record's separate license was not unambiguously exposed in the retrieved metadata, so confirm it with the publisher before redistribution. GISSchools links the record externally and does not mirror weights.

Before using a result

1. Keep survey rights, acquisition dates and preprocessing with each experiment.
2. Separate locations before tiling to avoid train/test leakage.
3. Inspect failures and uncertain outputs alongside the original point cloud.
4. Preserve coordinate transforms and verify units before exporting a GIS layer.

Validation status: Source provenance and website delivery are checked. GISSchools has not run model inference, training or an accuracy benchmark for this record. Feature images and gallery diagrams are explanatory illustrations, not predictions.

Official references

• Original project
• Usage documentation
• Research paper
• Selected official checkpoint
• License and usage conditions