Overview
We collected our dataset over multiple recording session in a forest near Stein am Rhein, Switzerland. Overall, we collected data in multiple plots denoted by a letter indicating the main type of trees in the plots, as shown in the map.
We provide LiDAR point clouds recorded with a backpack-carried automotive LiDAR sensors used in robotics applications and additionally LiDAR point clouds from a UAV for the March recording sessions.
For the backpack data, we annotated all plots with semantics (tree, shrub, ground) and also provide instance annotations for trees. Additionally, we distinguish for each tree between the stem and foliage of the trees.
Additionally, we provide reference measurements of key forestry traits acquired by domain experts.
The following table provides an overview of the dataset contents, where we split the data into a training, validation, and test set.
| Plot | March 2023 | October 2023 | July 2024 | Split |
|---|---|---|---|---|
| C1 | G,A,L,R | G,L | G,L | Validation |
| D2 | G,A,L,R | G,L | G,L | Training |
| M1 | G,A,L,R | G | G | Training |
| M2 | G,A | G | G | Test |
| M3 | G,A,L,R | G | G | Training |
| M4 | G,A,L,R | G | G | Training |
| M5 | G,A,L,R | G | G | Training |
| "G" = Ground Data available, "A" = Aerial Data available, "L" = Labels available, "R" = Reference measurements available. | ||||
Data Format and Folder Structure
Besides the raw data as ROS bags, we also provide processed data, where all LiDAR scans have poses estimated via a LiDAR SLAM approach. For each plot, we have the following folder structure.
- exp07-m1
- ground_clouds
- cloud_1679401800_628520000.pcd
- cloud_1679401805_227596000.pcd
- cloud_1679401806_228290000.pcd
- ⋮
- aerial_clouds
- cloud_1679401800_628520000.pcd
- cloud_1679401805_227596000.pcd
- cloud_1679401806_228290000.pcd
- ⋮
- labels
- cloud_1679401800_628520000.label
- cloud_1679401805_227596000.label
- cloud_1679401806_228290000.label
- ⋮
- rosbags
- frontier_2023-03-21-12-29-58_0.bag
- frontier_2023-03-21-12-34-14_1.bag
- frontier_2023-03-21-12-38-29_2.bag
- ⋮
- poses.txt
Specifically, we have the following file formats:
- ground_clouds and aerial_clouds: Contains the under-canopy point clouds recorded with the mobile mapping system in the PCD format, where we have x,y,z coordinates in the local coordinate frame of the LiDAR sensor and the intensity.
- ground_labels: Contains for each point cloud in ground_clouds the semantic
annotations. The format follows the binary SemanticKITTI format, where the lower 16-bit contain the
label and fine-grained label and the upper 16-bits contain the instance id. More specifically, we use an
32-bit unsigned int for each point, where bits 31-16 correspond to the instance id, bits 15-8 are
the fine-grained semantics, bits 7-0 are the semantic label:
- poses.txt: Contains the per scan poses, where poses are associated via the timestamp to individual scans. Each line of the csv file contains the following columns: counter,sec,nsec,x,y,z,qx,qy,qz,qw. Here, x,y,z correspond to the translation \(\mathbf{t}\in \mathbb{R}^3\) and qx,qy,qz,qw to a quaternion representing the rotational part \(\mathbf{R} \in \mathbb{R}^{3\times 3}\) that can be used to transform every point \(\textbf{p}_{\text{local}}\) into a point \(\textbf{p}_{\text{global}}\) in the world coordinate frame: $$\mathbf{p}_{\text{global}} = \mathbf{R}\, \mathbf{p}_{\text{local}} + \mathbf{t}$$
Development Kit and Baseline Code
For more convenient usage of the dataset, we provide a development kit that provides code for reading and processing the data. Additionally, we provide the code of our baselines for panoptic forestry segmentation, which is available at:
Hidden Test Set Evaluation
We plan to provide a hidden test set evaluation via a CodaLab competition that provides unbiased and reproducible results evaluated on plot M2 data. (Coming soon)
Citation
If you use our dataset or the provided tools, it would be nice if you cite our paper (PDF):
@inproceedings{malladi2025icra,
author = {Meher V.R. Malladi and Nived Chebrolu and Irene Scacchetti and Luca Lobefaro and Tiziano Guadagnino and Beno\^{i}t Casseau
and Haedam Oh and Leonard Frei{\ss}muth and Markus Karppinen and Janine Schweier and Stefan Leutenegger and Jens Behley
and Cyrill Stachniss and Maurice Fallon},
title = {{DigiForests: A Longitudinal LiDAR Dataset for Forestry Robotics}},
journal = {Proc.~of the IEEE Intl. Conf. on Robotics \& Automation (ICRA)},
year = {2025},
}