SALT: A Flexible Semi-Automatic Labeling Tool for General LiDAR Point Clouds with Cross-Scene Adaptability and 4D Consistency

This repo is the official project repository of [SALT].

1. Overview

We introduce SALT, a flexible semi-automatic labeling tool for general LiDAR point clouds, featuring adaptive cross-scene and 4D consistency. SALT demonstrates exceptional zero-shot adaptability across various sensors, scenes, and motion conditions, greatly enhancing annotation efficiency.

2. Environment

2.1 Docker-based

This setup allows you to run SALT in an isolated and reproducible Docker container, without having to install dependencies manually.

Step 1: Run the Container

To run the container with access to your project files and enable GPU:

docker run -it \
  --gpus all \
  -e DISPLAY=$DISPLAY \
  -v /tmp/.X11-unix:/tmp/.X11-unix \
  -v ~/SALT:/workspace/SALT \
  -w /workspace/SALT \
  cytt/salt:latest /bin/bash

Step 2: Run SALT Inside Docker

Once inside the container:

cd ~/SALT/bin
./SALT

The GUI labeling tool should now launch inside Docker.

---

2.2 Non-Docker

The dependencies can be installed from the package manager (tested On Ubuntu 18.04):

sudo apt-get install g++ build-essential libeigen3-dev python3-pip \
 python3-dev cmake git libboost-all-dev qtbase5-dev libglew-dev libyaml-cpp-dev -y 

Then, ensure Anaconda is installed in your system. Create and activate a virtual environment named SALT.

conda create -n SALT python==3.10
conda activate SALT

SALT depends on Patchwork++, a real-time ground segmentation library. Install it by running:

cd 3rdparty
git clone https://github.com/url-kaist/patchwork-plusplus.git && cd patchwork-plusplus
make pyinstall
cd ..

SAM 2 is required for the segmentation module. The code requires torch>=2.5.1 and torchvision>=0.20.1. Please follow the instructions here to install both PyTorch and TorchVision dependencies. You can install SAM 2 on a GPU machine using:

cd sam2
pip install -e .
cd checkpoints && ./download_ckpts.sh
cd ../../..

or individually download checkpoint from SAM 2.

Install SALT dependencies and download Data Alignment Model checkpoints.

pip install -r requirements.txt
cd ..

Meanwhile, please manually download the Data_Alignment_Model and place it in the ~/SALT/SALT/ directory.

The project folder structure should look like this:

SALT folder
├── 3rdparty
├── SALT 
   ├── best_model.pth    -- Data_Alignment_Model
   ├── ...
├── src   
└── ...

Install system-level dependencies and build the C++ backend of the labeler tool:

mkdir build && cd build
cmake ..
make -j5

Now the project root directory (e.g. ~/SALT) should contain a bin directory containing the labeler.In the bin directory, just run ./SALT to start the labeling tool.

3. User Manual

3.1 Automatic Segmentation of Whole sequence

After loading the raw point cloud sequence data, the user only needs to click the "SALT" button once to obtain the pre-segmentation results for the entire sequence. After clicking the "SALT" button, users can modify the config file according to their own data characteristics. Once the progress bar (zero-shot segmentation step) is complete, the pre-segmentation results are automatically saved for subsequent semantic and instance labeling. The pre-segmentation results are also automatically displayed in the user interface with different colors.

3.2 Manual Assignment and Refinement for Semantic Annotation

The user is free to define as many semantic classes appear in the sequence. Users can assign custom semantic labels to the pre-segmentation results based on their needs. By simply clicking on a predefined color button representing a specific semantic category and then selecting a point cloud with a particular ID, all points with that ID will be assigned the chosen label and updated to the corresponding color. This operation is as intuitive and effortless as a coloring game. Please note that the colors used to display the pre-segmentation results are designed to avoid overlapping with user-defined semantic label colors. If users are not satisfied with the pre-annotated results, they can modify them using the polygon tool. Inherited from LABELER, SALT supports the option to hide other classes, making manual annotation adjustments more convenient.

3.3 Auto Ordering and Manual Refinement for Instance Annotation

Once users are satisfied with the semantic labeling results, they can simply click the "Auto Instance" button to automatically assign instance IDs to all semantic categories. Users can then further refine the results by splitting or merging instance IDs within each category.

4. Quick Demo

We provide a simple demo dataset to quickly test the functionality of SALT.

You can download the demo data from here.This is a sample subset from the SemanticKITTI.

After downloading and extracting the data, you will find a config.yaml file inside the demo directory. You can use this configuration to quickly get started by simply updating a few path fields in the config file.

Step 1: Configure config.yaml

Open the file at:

<SALT_path>/SALT/config.yaml

Update the following fields to point to your local environment:

• conda_sh_path: Path to your conda.sh, e.g., /home/user/anaconda3/etc/profile.d/conda.sh
• data_dir: Path to the demo data folder, e.g., /home/user/demo_data
• cache_dir: Path to the cache output folder, e.g., /home/user/demo_data/cache
• sam2_checkpoint: Path to the SAM2 checkpoint, e.g., /home/user/SALT/3rdparty/sam2/checkpoints/sam2.1_hiera_large.pt
• model_cfg: Path to the SAM2 model configuration file, e.g., //home/user/SALT/3rdparty/sam2/sam2/configs/sam2.1/sam2.1_hiera_l.yaml
• Computing device configuration: depends on the machine you are using.

You can refer to the provided config.yaml as a template.

Step 2: Launch the GUI Tool

Once your paths are configured, simply launch the GUI and click the "SALT" button in the top toolbar to start labeling.

cd bin
./SALT

This will load the demo dataset using your customized config.yaml and bring up the annotation interface.

Enjoy labeling!

5. Acknowledgements

We would like to thank all the pioneers SemanticKITTI_LABLER, SAM2.

6. Citation

If your like our projects, please cite us and give this repo a star.

@article{wang2025salt,
  title={SALT: A Flexible Semi-Automatic Labeling Tool for General LiDAR Point Clouds with Cross-Scene Adaptability and 4D Consistency},
  author={Wang, Yanbo and Chen, Yongtao and Cao, Chuan and Deng, Tianchen and Zhao, Wentao and Wang, Jingchuan and Chen, Weidong},
  journal={arXiv preprint arXiv:2503.23980},
  year={2025}
}