🗂 GridNet-HD dataset

1. Introduction

This dataset was developed for 3D semantic segmentation tasks using both images and 3D point clouds specialized on electrical infrastructure. Grid (electrical) Network at High Density and High Resolution represents the first Image+LiDAR dataset accurately co-referenced in the electrical infrastructure domain. This dataset is associated with a public leaderboard hosted on Hugging Face Spaces, available at: leaderboard.

The dataset is associated with the following paper:

Title: GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure

Authors: Masked for instance

Conference: Submitted to NeurIPS 2025

This repository hosts the official data splits and resources used in the experiments reported in the paper.

---

2. Dataset Structure

This dataset consists of 36 geographic zones, each represented by a folder named after its area code (e.g. t1z4, t1z5a, etc.).

Each zone contains aligned multimodal data (images, segmentation masks, LiDAR point cloud, and camera parameters), enabling high-precision image-to-3D projection for multimodal fusion 3D semantic segmentation task.

A split.json file at the root of the dataset defines the official train/test partition of the zones.

To ensure fair evaluation on the official test set, ground truth annotations are not provided for either the images or the LiDAR point clouds. Instead, participants must submit their predictions to the leaderboard, where the official metrics (mIoU) are automatically computed against the hidden labels.

📁 Folder layout

dataset-root/
├── t1z5b/
│   ├── images/           # RGB images (.JPG)
│   ├── masks/            # Semantic segmentation masks (.png, single-channel label)
│   ├── lidar/            # LiDAR point cloud (.las format with field "ground_truth")
│   └── pose/             # Camera poses and intrinsics (text files)
├── t1z6a/
│   ├── ...
├── ...
├── split.json            # JSON file specifying the train/test split
└── README.md

🧾 Contents per zone

Inside each zone folder, you will find:

• 📷 images/

  • High-resolution RGB images (.JPG)
  • Captured from a UAV

• 🏷️ masks/

  • One .png mask per image, same filename as image
  • Label-encoded masks (1 channel)

• 🌍 lidar/

  • Single .las file for the entire zone captured from a UAV
  • Contains 3D point cloud data at high denisty with semantic ground_truth labels (stored in field named "ground_truth")

• 📌 pose/

  • camera_pose.txt: Camera positions and orientations per image (using Metashape Agisoft convention, more details in paper)
  • camera_calibration.xml: Camera calibration parameters (using Metashape Agisoft calibration model)

---

3. Class Grouping

Original classes have been grouped into 12 semantic groups as follows:

Group ID	Original Classes	Description
0	0,1,2,3,4	Pylon
1	5	Conductor cable
2	6,7	Structural cable
3	8,9,10,11	Insulator
4	14	High vegetation
5	15	Low vegetation
6	16	Herbaceous vegetation
7	17,18	Rock, gravel, soil
8	19	Impervious soil (Road)
9	20	Water
10	21	Building
255	12,13,255	Unassigned-Unlabeled

If interested the original classes are described in the Appendices of the paper.

📝 Note: group (12,13,255) is ignored during official evaluations.

---

4. Dataset Splits

The dataset is split into two parts:

• Train (~70% of LiDAR points)
• Test (~30% of LiDAR points)

The splits were carefully constructed to guarantee:

• Full coverage of all semantic groups (except the ignored group)
• No project overlap between train and test
• Balanced distribution in terms of class representation

Project assignments are listed in split.json with a proposal of split train/val.

Note that the test set give only the LiDAR without labels (without ground_truth field) and without mask labeled for images, this label part is keep by us in private mode for leaderboard management. To submit results on test set and obtain mIoU score on leaderboard, please follow instructions here: leaderboard on the remap classes presented below.

---

5. Dataset Statistics

📈 Class Distribution

The table below summarizes the number of points per semantic group across the train and test splits, including the total number of points, the proportion of each class present in the test set (% test/total), and the relative class distributions within each split (Distribution classes in train/test set (%)).

Group ID	Train Points	Test Points	Total points	% test/total	Distribution classes in train set (%)	Distribution classes in test set (%)
0	11'490'104	3'859'573	15'349'677	25.1	0.7	0.5
1	7'273'270	3'223'720	10'496'990	30.7	0.4	0.4
2	1'811'422	903'089	2'714'511	33.3	0.1	0.1
3	821'712	230'219	1'051'931	21.9	0.05	0.03
4	278'527'781	135'808'699	414'336'480	32.8	16.5	17.9
5	78'101'152	37'886'731	115'987'883	32.7	4.6	5.0
6	1'155'217'319	461'212'378	1'616'429'697	28.5	68.4	60.7
7	135'026'058	99'817'139	234'843'197	42.5	8.0	13.1
8	13'205'411	12'945'414	26'150'825	49.5	0.8	1.7
9	1'807'216	1'227'892	3'035'108	40.5	0.1	0.2
10	6'259'260	2'107'391	8'366'651	25.2	0.4	0.3
TOTAL	1'689'540'705	759'222'245	2'448'762'950	31.0	100	100

The same table summarizes the same features as above for the proposed split train/val:

Group ID	Train Points	Val Points	Total points	% val/total	Distribution classes in train set (%)	Distribution classes in val set (%)
0	8'643'791	2'846'313	11'490'104	24.8	0.7	0.7
1	5'782'668	1'490'602	7'273'270	20.5	0.4	0.4
2	1'370'331	441'091	1'811'422	24.4	0.1	0.1
3	625'937	195'775	821'712	23.8	0.05	0.05
4	160'763'512	117'764'269	278'527'781	42.3	12.4	29.7
5	43'442'079	34'659'073	78'101'152	44.4	3.4	8.7
6	968'689'542	186'527'777	1'155'217'319	16.1	74.9	47.0
7	87'621'550	47'404'508	135'026'058	35.1	6.8	11.9
8	10'420'302	2'785'109	13'205'411	21.1	0.8	0.7
9	310'240	1'496'976	1'807'216	82.8	0.02	0.4
10	4'793'225	1'466'035	6'259'260	23.4	0.4	0.4
TOTAL	1'292'463'177	397'077'528	1'689'540'705	23.5	100	100

📈 Class Distribution Visualisation

---

6. How to Use

Download via Hugging Face Hub

⚠️ Warning: This dataset is large, the full download size is approximately 170 GB. Make sure you have sufficient disk space and a stable internet connection before downloading.

To download the full dataset, please don't use the function datasets.load_dataset() from huggingface, this parquet version of the dataset is automatically done by huggingface but not adapted for this type of dataset (>5GB).

Use instead:

from huggingface_hub import snapshot_download
local_dir = snapshot_download(
    repo_id="heig-vd-geo/GridNet-HD",
    repo_type="dataset",
    local_dir="GridNet-HD" # where to replicate the file tree
)

---

7. Running baselines

Please follow instructions on dedicated git repository for models running on this dataset:

• Baseline based on image segmentation and reprojection into LiDAR: ImageVote baseline
• Baseline based on LiDAR 3D segmentation directly using Superpoint Trasnformer (SPT): SPT baseline
• Baseline based on late fusion between softmax logits from SPT and ImageVote: LateFusionMLP baseline

Results are visible here with the 3 different baselines:

Baseline	ImageVote baseline	ImageVote baseline	SPT baseline	SPT baseline	Late fusion MLP	Late fusion MLP
Class	IoU (Validation set) (%)	IoU (Test set) (%)	IoU (Validation set) (%)	IoU (Test set) (%)	IoU (Validation set) (%)	IoU (Test set) (%)
Pylon	86.43	85.09	93.92	92.75
Conductor cable	58.22	64.82	89.36	91.05
Structural cable	48.84	45.06	68.44	70.51
Insulator	81.16	71.07	85.79	80.60
High vegetation	90.89	83.86	87.60	85.15
Low vegetation	71.89	63.43	57.95	55.91
Herbaceous vegetation	93.87	84.45	91.40	84.64
Rock, gravel, soil	91.40	38.62	82.42	40.63
Impervious soil (Road)	85.84	80.69	74.12	73.57
Water	44.14	74.87	50.94	3.69
Building	86.76	68.09	73.97	57.38
Mean IoU (mIoU)	76.31	69.10	77.81	66.90

---

8. License and Citation

This dataset is released under the CC-BY-4.0 license.

If you use this dataset, please cite the following paper:

GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure
Masked Authors
Submitted to NeurIPS 2025.