This is the official repo of HARP-NeXt, a High-Speed and Accurate Range-Point Fusion Network for 3D LiDAR Semantic Segmentation. We first propose a novel pre-processing methodology that significantly reduces computational overhead. Then, we design the Conv-SE-NeXt feature extraction block to efficiently capture representations without deep layer stacking per network stage. We also employ a multi-scale range-point fusion backbone that leverages information at multiple abstraction levels to preserve essential geometric details, thereby enhancing accuracy.

Fast and Accurate 3D Scene Semantic Segmentation on nuScenes

• [2025.10] - We provide trained network weights on nuScenes and SemanticKITTI benchmarks. The checkpoints are available here.
• [2025.10] - Our paper is available on arXiv, and the code is publicly available.
• [2025.06] - Our paper was accepted by IROS 2025.

HARP-NeXt achieves high accuracy in real-time on the RTX 4090 GPU at 100 FPS, while delivering near real-time performance on the Jetson AGX Orin embedded system, breaking performance trends.

mIoU vs. Runtime on the NVIDIA RTX 4090 GPU	mIoU vs. Runtime on the NVIDIA Jetson AGX Orin

We recommend using conda with the conda-forge channel.

We recommend Python 3.10–3.12 (tested with 3.12)

conda create -n harpnext -c conda-forge python=3.12
conda activate harpnext

We use PyTorch v2.5.0 with CUDA 11.8:

pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu118

conda install -c conda-forge numpy tqdm tensorboard
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.5.0+cu118.html
pip install pyyaml seaborn opencv-python

We recommend using venv virtual environment and installing PyTorch (for JetPack) following NVIDIA's documentation. The following installation is tested on an NVIDIA Jetson AGX Orin Developer Kit with a JetPack 6.0 (L4T R36.2 / R36.3).

We recommend creating a Python 3 venv:

python3 -m venv harpnext
source harpnext/bin/activate

We use PyTorch v2.3.0 with CUDA 12.4:

wget https://nvidia.box.com/shared/static/zvultzsmd4iuheykxy17s4l2n91ylpl8.whl -O torch-2.3.0-cp310-cp310-linux_aarch64.whl
sudo apt-get install python3-pip libopenblas-base libopenmpi-dev libomp-dev
pip3 install 'Cython<3'
pip3 install numpy torch-2.3.0-cp310-cp310-linux_aarch64.whl

pip3 install tqdm tensorboard
pip3 install torch-scatter
pip3 install pyyaml seaborn opencv-python

You can download the nuScenes dataset from the official website with the following structure:

nuscenes
├── lidarseg
|   ├── v1.0-mini
|   ├── v1.0-test
|   ├── v1.0-trainval
├── panoptic
├── samples
├── v1.0-mini
├── v1.0-test
├── v1.0-trainval

You can download the SemanticKITTI dataset from the official website with the following structure:

semantickitti
├── sequences
│   ├── 00
│   │   ├── labels
│   │   ├── velodyne
│   ├── 01
│   ├── ..
│   ├── 21

You can download the pretrained weights on nuScenes and SemanticKITTI from here and save them in the ./pretrained/ directory.

You can evaluate HARP-NeXt's pretrained model on nuScenes as follows:

python main.py \
--net harpnext \
--dataset nuscenes \
--path_dataset /path/to/nuscenes \
--mainconfig ./configs/main/main-config.yaml \
--netconfig ./configs/net/harpnext-nuscenes.yaml \
--log_path ./pretrained/harpnext-nuscenes-32x480 \
--gpu 0 \
--seed 0 \
--fp16 \
--restart \
--eval

You can evaluate HARP-NeXt's pretrained model on SemanticKITTI as follows:

python main.py \
--net harpnext \
--dataset semantic_kitti \
--path_dataset /path/to/SemanticKITTI \
--mainconfig ./configs/main/main-config.yaml \
--netconfig ./configs/net/harpnext-semantickitti.yaml \
--log_path ./pretrained/harpnext-cutmix-semantickitti-64x512 \
--gpu 0 \
--fp16 \
--restart \
--eval

You can choose whether to execute the pre-processing on CPU or GPU by setting the option in the corresponding netconfig harpnext-nuscenes.yaml and harpnext-semantickitti.yaml files.

You can retrain the harpnext-nuscenes-32x480 model on nuScenes as follows:

python main.py \
--net harpnext \
--dataset nuscenes \
--path_dataset /path/to/nuscenes \
--mainconfig ./configs/main/main-config.yaml \
--netconfig ./configs/net/harpnext-nuscenes.yaml \
--log_path ./logs/harpnext-nuscenes-32x480-retrain \
--gpu 0 \
--seed 0 \
--fp16

You can retrain the harpnext-cutmix-semantickitti-64x512 model on SemanticKITTI as follows:

python main.py \
--net harpnext \
--dataset semantic_kitti \
--path_dataset /path/to/SemanticKITTI \
--mainconfig ./configs/main/main-config.yaml \
--netconfig ./configs/net/harpnext-semantickitti.yaml \
--log_path ./logs/harpnext-cutmix-semantickitti-64x512-retrain \
--gpu 0 \
--fp16

You can enable or disable the Instance CutMix and PolarMix augmentations, during the training only on SemanticKITTI, by setting the instance_cutmix flag in the main-config.yaml file (True to enable, False to disable). The extracted instances are saved in /tmp/semantic_kitti_instances/. Always set to False when evaluating the models.

This work is under the Apache 2.0 license.

Kindly consider citing our paper if you find this work helpful:

@article{abouhaidar2025harpnext,
  title={HARP-NeXt: High-Speed and Accurate Range-Point Fusion Network for 3D LiDAR Semantic Segmentation},
  author={Abou Haidar, Samir and Chariot, Alexandre and Darouich, Mehdi and Joly, Cyril and Deschaud, Jean-Emmanuel},
  journal={arXiv preprint arXiv:2510.06876},
  year={2025}
}

We acknowledge and thank the following public resources used in this work:

• SemanticKITTI (CC BY-NC-SA 4.0)
• SemanticKITTI-API (MIT)
• nuScenes (CC BY-NC-SA 4.0)
• nuScenes-devkit (Apache 2.0)
• SalsaNext (MIT)
• CENet (MIT)
• MinkowskiNet
• SPVNAS (MIT)
• Cylinder3D (Apache 2.0)
• FRNet (Apache 2.0)
• WaffleIron (Apache 2.0)
• 2DPASS (MIT)
• PTv3 (MIT)
• PointNeXt (MIT)
• RPVNet
• Deep Residual Networks (MIT)
• ConvNeXt (MIT)
• MobileNetv3
• Squeeze-and-Excitation
• LovaszSoftmax (MIT)
• Boundary Loss (MIT)
• PolarMix (MIT)
• LaserMix (Apache 2.0)