HORTO-3DLM Dataset

The HORTO-3DLM Dataset contains 3D LiDAR and GNSS/localization data for the purpose of 3D LiDAR-based place recognition, localization and mapping in horticultural environments.

Data Download |

Content

Updates
Dataset
Citation

1. Updates

25/02/2024 HORTO-3DLM v2.0 added GTJ34 and ON22 sequences
1/12/2024 HORTO-3DLM v1.0 Uploaded

2. Dataset

Table Caption: Summary of all sequences

The Seq. column contains the sequence names. The M, Y, and C columns refer to the month, year, and country of recording, respectively. The total distance (Dist) of each sequence is measured in meters, while the scan size refers to the number of points in each scan.

Seq.	M	Y	C	Nº Scans	Nº Rows	Dist. [m]	Scan Size	Plantation Type
ON22	Nov.	2022	UK	7974	4	514	48k	Apple (open)
OJ22	July	2022	UK	4361	3	206	50k	Apple (open)
OJ23	June	2023	UK	7229	3	459	46k	Cherry (open)
SJ23	June	2023	UK	6389	3	742	48k	Strawberry (polytunnels)
ON23	Nov.	2023	FR	3086	5	966	105k	Apple (open)
GTJ23	June	2023	PT	661	3	202	60k	Tomato (greenhouse)

Trajectories with segment labels and recording setups

Sequence ON23 was recorded in November 2023, in an orchard in Metz, France, with an 16-beam Ouster 3D LiDAR and an SBG GNSS/INS system (without RTK) mounted on a Clearpath Husky mobile platform. To address the low LiDAR resolution, the original scans were merged to increase point density, resulting in sub-maps with approximately 100k points per sub-map. This operation reduced the original sequence from 25836 scans to 3086 sub-maps in total.

Sequence GTJ23 was recorded in June of 2023, in a tomato plantation within a greenhouse, in Coimbra, Portugal, with a 64-beam Ouster 3D LiDAR mounted on a Clearpath Jackal mobile platform. Due to signal interference caused by the greenhouse structure, the GNSS signal was unreliable. Therefore, the ground-truth positions were computed using a SLAM approach.

V1.0:

These sequences were recorded in England, UK, using a Clearpath Husky mobile robot equipped with a Velodyne VLP32 3D LiDAR (10Hz) and a ZED-F9P RTK-GPS (5Hz).

OJ23: orchards,
OJ22: orchards,
SJ23: strawberries within polytunnels with a table-top growing system.
ON22: orchards

3D Maps

Structure

HORTO-3DLM
├── GT23 
├── OJ22
├── OJ23
├── ON22
├── ON23
└── SJ23
    └── extracted
        ├── eval
        |   └──ground_truth_loop_range_10m.pkl 
        ├── point_cloud
        |   ├── 0000000.bin
        |   ├── 0000001.bin
        |   ├── ...
        |   └── xxxxxxx.bin
        ├── triplet
        |   └── ground_truth_ar0.5m_nr10m_pr2m.pkl
        ├── extracted_info.txt
        ├── gps.txt
        ├── gps.kmz
        ├── gps_timestamp.txt
        ├── point_cloud_timestamp.txt
        ├── point_row_labels.pkl
        ├── positions.txt
        ├── positions_timestamp.txt
        ├── static_tf.txt
        ├── tf_poses.txt
        └── tf_poses_timestamp.txt

Content:

eval/ground_truth_loop_range_10m.pkl: pickle file with the ground truth data for evaluation(testing), containing anchors/positives indices for 10m range loops.
point_cloud/: Folder with the point clouds.
triplet/ground_truth_ar0.5m_nr10m_pr2m.pkl pickle file with the ground truth data for training, containing anchors, positives and negatives. The 0.5m between anchors; the positives were selected from within a 2m range; and negatives were selected from outside a 10m range.
extracted_info.txt: some info from the extracted data;
gps.txt: GNSS data in UTM reference system.
gps.kmz: GNSS data for visualization on maps.
gps_timestamp.txt: timestamps for GNSS data.
point_cloud_timestamp.txt: timestamps for point cloud. row index 00 -> points cloud 00000.bin.
point_row_labels.pkl: segment labels for each position data.
positions.txt: position data with (x,y,z), used in place recognition for ground truth.
positions_timestamp.txt: timestamps for position data
static_tf.txt: static tf between base_link and LiDAR
tf_poses.txt: pose data in the format of transformation matrix flattened to fit one file row.
tf_poses_timestamp.txt: timestamp of the tf_pose.txt data.

Citation:

@article{barros2023orchnet,
    title={ORCHNet: A Robust Global Feature Aggregation approach for 3D LiDAR-based Place recognition in Orchards},
    author={Barros, T and Garrote, L and Conde, P and Coombes, MJ and Liu, C and Premebida, C and Nunes, UJ},
    journal={arXiv preprint arXiv:2303.00477},
    year={2023}
}

TO-DO

Add information
Add Sensor TFs
Add Dataset structure

Cybonic/HORTO-3DLM