PointDenseBEV: Dense Semantic Bird’s-Eye-View Map Generation

This is the official implementation of the IROS2025 paper Dense Semantic Bird-Eye-View Map Generation from Sparse LiDAR Point Clouds via Distribution-aware Feature Fusion.

Click the image above to watch the demonstration video.

📖 Overview

We propose PointDenseBEV, an end-to-end, distribution-aware feature fusion framework. It takes sparse LiDAR point clouds as input and directly generates dense semantic Bird's-Eye-View (BEV) maps. Spatial geometric information and temporal context are embedded as auxiliary semantic cues within the BEV grid representation to improve semantic density.

Extensive experiments on the SemanticKITTI dataset demonstrate that our method achieves competitive performance compared with existing approaches.

Note: This project will continue to be refined in the near future, including environment mirroring, data preprocessing, and weighting results.

🔗 Resources

• Base Framework: Built on top of OpenPCDet.
• Project Files & Dataset: Download via OneDrive

---

🚀 Getting Started

1. Environment Setup (Docker)

We strongly recommend using Docker to set up the environment. You can download the Docker environment package (segkitti_250114.tar) from the provided OneDrive link and build the image following the standard Docker guide.

Run the following command to start the container:

docker run --name seg_kitti \
    --shm-size=128g \
    --gpus all \
    -dit \
    -p 9800:9800 \
    -v /path/to/your/Datasets:/home/Datasets \
    -v /path/to/your/Projects:/home/Projects \
    c3a505511a3f \
    /bin/bash

(Replace /path/to/your/... and the image ID c3a505511a3f with your local paths and actual loaded image ID if necessary).

2. Installation

Since this project is built upon OpenPCDet, it needs to be installed before running.

# Clone the repository (if not already done) and enter the project root
cd <PROJECT_ROOT>

# Install the project
python setup.py develop

---

📂 Data Preparation

Download

1. Download the density maps (kitti_vis_8) and ground-truth labels (dense_label, sparse_label) from the provided OneDrive link.
2. Obtain the raw point clouds from the official KITTI website.

Directory Structure

After decompression, organize your dataset directory as follows:

<PROJECT_ROOT>
└── datasets/
    └── dense_kitti/
        ├── dense_label/        # Dense-mode ground truth
        │   ├── 00/
        │   │   └── *.bin
        │   ├── ...
        │   └── 10/
        ├── kitti_odo/          # Raw point clouds
        │   ├── 00/
        │   ├── ...
        │   └── 10/
        ├── kitti_vis_8/        # Visibility map files
        │   └── ...
        └── sparse_label/
            └── ...

---

🏃 Training and Evaluation

Training

Run the training script to train the model.

sh tools/train.sh

Configuration Options: You can modify the model and dataset configurations in the following files:

• tools/cfgs/kitti_models/densepillar.yaml
• tools/cfgs/dataset_configs/test_pillarnet_128.yaml

Example configuration modifications:

1. Update dataset paths (kitti_dataset.yaml):

DATA_LOAD_PATH: '/home/Projects/DensePillar/datasets/dense_kitti'
DATA_TRAIN_PATH: '/home/Projects/DensePillar/datasets/dense_kitti/train_sample.pkl'
DATA_TEST_PATH: '/home/Projects/DensePillar/datasets/dense_kitti/test_sample.pkl'

2. Enable multi-frame point cloud data:

POINT_MERGE_HISFRAME: 0

3. Adjust voxel numbers (test_pillarnet_128.yaml):

MAX_POINTS_PER_VOXEL: 10
MAX_NUMBER_OF_VOXELS: {
    'train': 50000, 
    'test': 50000
}

Evaluation

We provide pre-trained weights for single-frame LiDAR point cloud input.

1. Download checkpoint_epoch_11.pth from the OneDrive link.
2. Place it in the following directory: logs/model/kitti_models/test_pillarnet_128/pretrain/ckpt/
3. Evaluate the trained model (make sure to set the ckpt_dir path properly inside the script):

sh tools/test.sh

Visualization: If you want to visualize the evaluation results, uncomment the visualization code block in tools/eval_utils/eval_utils.py.

---

📚 Citation

If you find our work helpful, please cite:

@INPROCEEDINGS{li2025dense,
  author={Jinsong Li and Kunyu Peng and Yuxiang Sun},
  booktitle={2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={Dense Semantic Bird-Eye-View Map Generation from Sparse LiDAR Point Clouds via Distribution-aware Feature Fusion}, 
  year={2025},
  volume={},
  number={},
  pages={4123-4129},
  doi={10.1109/IROS60139.2025.11246382}}

---

🙏 Acknowledgements

Special thanks to:

• OpenPCDet for the highly extensible 3D object detection codebase.
• MASS (Multi-Attentional Semantic Segmentation of LiDAR Data for Dense Top-View Understanding) for their inspiring contributions to the community.