| --- |
| library_name: pytorch |
| tags: |
| - point-cloud |
| - point-cloud-backbone |
| - self-supervised-learning |
| - lidar |
| - forest |
| - pytorch |
| authors: |
| - Yuanwen Yue |
| - Stefano Puliti |
| - Damien Robert |
| - Atakan Topaloğlu |
| - Binbin Xiang |
| - Maciej Wielgosz |
| - Jan Dirk Wegner |
| - Rasmus Astrup |
| - Christian Rupprecht |
| - Konrad Schindler |
| --- |
| |
| This repository contains model weights for the paper **Toward A Foundation Model for Forest Point Clouds**. |
|
|
| ForPT pretrains a [LitePT](https://github.com/prs-eth/LitePT) backbone on a large corpus of unlabeled airborne (ALS), UAV (ULS) and mobile (MLS) laser scans covering diverse forest ecosystems. It is evaluated on four forestry tasks: semantic segmentation, instance segmentation, tree species classification and tree age regression. Self-supervised pretraining accelerates convergence, improves performance when annotations are scarce, and transfers better than task-specific supervised pretraining. |
|
|
| ## Paper & Resources |
| - **Arxiv:** [https://arxiv.org/abs/2609.24787](https://arxiv.org/abs/2609.24787) |
| - **Project Page:** [https://prs-eth.github.io/ForPT](https://prs-eth.github.io/ForPT) |
| - **Codebase:** [https://github.com/prs-eth/ForPT](https://github.com/prs-eth/ForPT) |
|
|
| ## Models |
| We release the pretrained backbone and the downstream models for the benchmarks reported in our paper. |
|
|
| ### Pretrained backbone |
| | Model | Params | Pretraining data | Config | Checkpoint | |
| |:-|-:|:-:|:-:|:-:| |
| | ForPT (LitePT-S) | 12.4M | ForPT corpus | [link](https://github.com/prs-eth/ForPT/blob/main/configs/00_ssl_pretrain/pretrain-forpt.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/00_ssl_pretrain/pretrain-forpt/model/epoch_30.pth) | |
|
|
| ### Downstream tasks (100% labels) |
| | Task | Benchmark | Model | Config | Checkpoint | |
| |:-|:-:|:-|:-:|:-:| |
| | Semantic segmentation | FOR-instanceV2 | LitePT-scratch | [link](https://github.com/prs-eth/ForPT/blob/main/configs/01_forinstancev2_sem/semseg-litepts-scratch-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/01_forinstancev2_sem/semseg-litepts-scratch-p100/model/model_best.pth) | |
| | | | ForPT-linear probing | [link](https://github.com/prs-eth/ForPT/blob/main/configs/01_forinstancev2_sem/semseg-forpt-lin-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/01_forinstancev2_sem/semseg-forpt-lin-p100/model/model_best.pth) | |
| | | | ForPT-decoder probing | [link](https://github.com/prs-eth/ForPT/blob/main/configs/01_forinstancev2_sem/semseg-forpt-dec-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/01_forinstancev2_sem/semseg-forpt-dec-p100/model/model_best.pth) | |
| | | | ForPT-finetune | [link](https://github.com/prs-eth/ForPT/blob/main/configs/01_forinstancev2_sem/semseg-forpt-ft-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/01_forinstancev2_sem/semseg-forpt-ft-p100/model/model_best.pth) | |
| | Instance segmentation | FOR-instanceV2 | Coming soon | - | - | |
| | Species classification | FOR-species20K | LitePT-scratch | [link](https://github.com/prs-eth/ForPT/blob/main/configs/03_forspecies20k/cls-litepts-scratch-allpts.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/03_forspecies20k/cls-litepts-scratch-allpts/model/model_last.pth) | |
| | | | ForPT-linear probing | [link](https://github.com/prs-eth/ForPT/blob/main/configs/03_forspecies20k/cls-forpt-lin-allpts.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/03_forspecies20k/cls-forpt-lin-allpts/model/model_last.pth) | |
| | | | ForPT-finetune | [link](https://github.com/prs-eth/ForPT/blob/main/configs/03_forspecies20k/cls-forpt-ft-allpts.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/03_forspecies20k/cls-forpt-ft-allpts/model/model_last.pth) | |
| | Age regression | FORage | LitePT-scratch | [link](https://github.com/prs-eth/ForPT/blob/main/configs/04_forage/age-litepts-scratch-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/04_forage/age-litepts-scratch-p100/model/model_best.pth) | |
| | | | ForPT-linear probing | [link](https://github.com/prs-eth/ForPT/blob/main/configs/04_forage/age-forpt-lin-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/04_forage/age-forpt-lin-p100/model/model_best.pth) | |
| | | | ForPT-finetune | [link](https://github.com/prs-eth/ForPT/blob/main/configs/04_forage/age-forpt-ft-p100.py) | [Download](https://huggingface.co/prs-eth/ForPT/resolve/main/04_forage/age-forpt-ft-p100/model/model_best.pth) | |
|
|
| To download all checkpoints into the layout expected by the configs, run from the root of the [ForPT codebase](https://github.com/prs-eth/ForPT): |
| ```shell |
| hf download prs-eth/ForPT --local-dir weights --exclude ".gitattributes" --exclude "README.md" |
| ``` |
|
|
| ## Citation |
|
|
| ``` |
| @article{yue2026forpt, |
| title={Toward a foundation model for forest point clouds}, |
| author={Yue, Yuanwen and Puliti, Stefano and Robert, Damien and Topalo{\u{g}}lu, Atakan and Xiang, Binbin and Wielgosz, Maciej and Wegner, Jan Dirk and Astrup, Rasmus and Rupprecht, Christian and Schindler, Konrad}, |
| journal={arXiv preprint arXiv:2609.24787}, |
| year={2026} |
| } |
| ``` |
|
|
