Data and model weights for completing 3D point clouds of individual trees using deep learning

This repository contains data associated with the study "Completing 3D point clouds of individual trees using deep learning"
and accompanies the code available at https://github.com/alBrnd/treePoinTr

DATA & FILE OVERVIEW

treePoinTr_rawdata

TheGrove_obj: 3D mesh files in .obj format, created using the Blender add-on "The Grove"
TheGrove_pointclouds: point clouds derived from TheGrove_obj by random sampling on mesh surface
SapTreeGen_obj: 3D mesh files in .obj format, created using the Blender add-on "Sapling Tree Gen"
SapTreeGen_pointclouds: point clouds derived from SapTreeGen_obj by random sampling on mesh surface
HeliosSim: HELIOS++ simulations on TheGrove objects. 12 scan positions (legs) per tree object, poit clouds in .xyz format.
realTLS_BLK360 trees: 12 individual tree leaf-off point clouds, acquired with a Leica BLK360 TLS. Branches < 5cm were removed.
realTLS_Riegl-VZ400_segments: Segments from forest point clouds acquired with a Riegl-VZ400i TLS. Areas were manually segmented for completeness.

treePoinTr_trainedModels

PoinTr models fine-tuned on the datasets with the same name.
The file ckpt-best.pth contains the best performing model weights and should be used for inference.

The samples generated from the raw data to train the models can be made available upon request (135 GB), but generating your own training samples using the provided scripts is recommended (make_samples_for_treePoinTr.ipynb).