heig-vd-geo/GridNet-HD
收藏资源简介：
--- license: cc-by-4.0 language: - en tags: - 3d - Image pretty_name: GridNet-HD size_categories: - 1B<n<10B task_categories: - feature-extraction --- # 🗂 GridNet-HD dataset ## 1. Introduction This dataset was developed for **3D semantic segmentation task** using both **images and 3D point clouds** specialized on electrical infrastructure. **Grid** (electrical) **Net**work at **H**igh **D**ensity 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](https://huggingface.co/spaces/heig-vd-geo/GridNet-HD-Leaderboard). The dataset is associated with the following paper: ``` @misc{gridnet-hd-dataset, title={GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure}, author={Antoine Carreaud and Shanci Li and Malo De Lacour and Digre Frinde and Jan Skaloud and Adrien Gressin}, year={2026}, eprint={2601.13052}, url={https://arxiv.org/abs/2601.13052}, } ``` This repository hosts the official data splits and resources used in the experiments reported in the paper. | Raw 3D | 3D segmented | Corresponding image | | --- | --- | --- | |  |  |  | --- ## 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](https://huggingface.co/spaces/heig-vd-geo/GridNet-HD-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 (5280x3956) (.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](https://huggingface.co/spaces/heig-vd-geo/GridNet-HD-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. 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](https://huggingface.co/heig-vd-geo/ImageVote_GridNet-HD_baseline) - Baseline based on LiDAR 3D segmentation directly using Superpoint Trasnformer (SPT): [SPT baseline](https://huggingface.co/heig-vd-geo/SPT_GridNet-HD_baseline) - Baseline based on late fusion between softmax logits from SPT and ImageVote: [LateFusionMLP baseline](https://huggingface.co/heig-vd-geo/LateFusionMLP_GridNet-HD_baseline) Results are visible here with the **best model** from 3 different baselines: | Baseline | ImageVote baseline | SPT baseline| Late fusion MLP| |---------------------------|-----------------------|---------------|------------------| | Class | IoU (Test set) (%)| IoU (Test set) (%)| IoU (Test set) (%)| | Pylon | 85.09 | 92.75 | **94.82** | | Conductor cable | 64.82 | 91.05 | **94.40** | | Structural cable | 45.06 | 70.51 | **82.52** | | Insulator | 71.07 | 80.60 | **86.98** | | High vegetation | 83.86 | **85.15** | 83.08 | | Low vegetation | **63.43** | 55.91 | 47.64 | | Herbaceous vegetation | 84.45 | **84.64** | 80.75 | | Rock, gravel, soil | 38.62 | 40.63 | **42.89** | | Impervious soil (Road) | **80.69** | 73.57 | 80.26 | | Water | **74.87** | 3.69 | 61.69 | | Building | **68.09** | 57.38 | 61.40 | | **Mean IoU (mIoU)** | 69.10 | 66.90 | **74.22** | --- ## 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: ``` @misc{gridnet-hd-dataset, title={GridNet-HD: A High-Resolution Multi-Modal Dataset for LiDAR-Image Fusion on Power Line Infrastructure}, author={Antoine Carreaud and Shanci Li and Malo De Lacour and Digre Frinde and Jan Skaloud and Adrien Gressin}, year={2026}, eprint={2601.13052}, url={https://arxiv.org/abs/2601.13052}, } ```
The GridNet-HD dataset is designed for 3D semantic segmentation tasks focusing on electrical infrastructure, utilizing both images and 3D point clouds. It comprises 36 geographic zones with aligned multimodal data, such as high-resolution RGB images, segmentation masks, LiDAR point clouds, and camera parameters. The dataset is split into train and test sets, ensuring full coverage of semantic groups and balanced class distribution. It is associated with a paper and accessible on Hugging Face Spaces, where participants can submit their predictions for evaluation. The dataset is released under the CC-BY-4.0 license and can be downloaded via the Hugging Face Hub. Baselines and instructions for running models on this dataset are also provided.
