ALMA pretrained models

ALMA: Airborne LiDAR Masked Autoencoding Framework for Self Supervised Representation Learning on 3DEP Point Cloud
Arnav Goel and Jinha Jung

Code and usage: https://github.com/gdslab/alma

Model weights: https://huggingface.co/gdslab/alma

Six artifacts are supplied: QL1/QL2 hierarchical encoders, full MAE reconstruction models, and complete final segmentation models. Download and verify using the code repository's manifest/downloader. SHA256SUMS lists the exact exported weight hashes. Optimizer states are omitted.

Regime	Crop	Pretraining best epoch	Segmentation best epoch
Hamilton QL1	100 ft	235	93
Indiana 3DEP QL2	200 ft	80 (stage 2)	65

Inputs are 20,480 XYZ points. Coordinates must be in feet. Pretraining uses bounding-box-centered XYZ divided by [50,50,25] (QL1) / [100,100,75] (QL2). Downstream uses mean-centered XY divided by half crop width and minimum-shifted Z in feet. Configurations are not interchangeable.

The segmentation decoder predicts LAS 2–6; other training labels are ignored, with geometry retained. Confidence-thresholded unclassified is post-processing. QL2 supervised evaluation uses a geometry-aware thinned QL1 proxy; native QL2 accuracy is not established. Models are regional and require validation on new data.

See the included model cards for provenance and limitations. The source checkpoints and every exported tensor were verified locally. The user reported all 12 release tests passing on 2026-09-30, including QL1/QL2 encoder and segmentation decoder GPU forwards. Fresh installation, full training, MAE CUDA loss/backward and real-tile execution have not been validated by this release audit.

Model weights are released under the MIT license; see WEIGHTS_LICENSE. Code retains its upstream MIT license and attribution; see LICENSE and NOTICE.