Learning to Drop Points for LiDAR Scan Synthesis (IROS 2021)

This repository provides the official PyTorch implementation of the following paper:

Learning to Drop Points for LiDAR Scan Synthesis
Kazuto Nakashima and Ryo Kurazume
In IROS 2021
[Project] [Paper]

Overview: We propose a noise-aware GAN for generative modeling of 3D LiDAR data on a projected 2D representation (aka spherical projection). Although the 2D representation has been adopted in many LiDAR processing tasks, generative modeling is non-trivial due to the discrete dropout noises caused by LiDAR’s lossy measurement. Our GAN can effectively learn the LiDAR data by representing such discrete data distribution as a composite of two modalities: an underlying complete depth and the corresponding reflective uncertainty.

Generation: The learned depth map (left), the measurability map for simulating the point-drops (center), and the final depth map (right).

Reconstruction: The trained generator can be used as a scene prior to restore LiDAR data.

Requirements

To setup with anaconda:

$ conda env create -f environment.yaml
$ conda activate dusty-gan

Datasets

Please follow the instruction for KITTI Odometry. This step produces the azimuth/elevation coordinates (angles.pt) aligned with the image grid, which are required for all the following steps.

Pretrained Models

Training

Please use train.py with dataset=, solver=, and model= options. The default configurations are defined under configs/ and can be overridden via a console (reference).

$ python train.py dataset=kitti_odometry solver=nsgan_eqlr model=dcgan_eqlr        # baseline
$ python train.py dataset=kitti_odometry solver=nsgan_eqlr model=dusty1_dcgan_eqlr # DUSty-I (ours)
$ python train.py dataset=kitti_odometry solver=nsgan_eqlr model=dusty2_dcgan_eqlr # DUSty-II (ours)

Each run creates a unique directory with saved weights and log file.

outputs/logs
└── dataset=<DATASET>
    └── model=<MODEL>
        └── solver=<SOLVER>
            └── <DATE>
                └── <TIME>
                    ├── .hydra
                    │   └── config.yaml # all configuration parameters
                    ├── models
                    │   ├── ...
                    │   └── checkpoint_0025000000.pth # trained weights
                    └── runs
                        └── <TENSORBOARD FILES> # logged losses, scores, and images

To monitor losses, scores, and images, run the following command to launch TensorBoard.

$ tensorboard --logdir outputs/logs

Evaluation

To evaluate synthesis performance, run:

$ python evaluate_synthesis.py --model-path $MODEL_PATH --config-path $CONFIG_PATH

MODEL_PATH and CONFIG_PATH are .pth file and the corresponding .hydra/config.yaml file, respectively. A relative tolerance to detect the point-drop can be changed via --tol option (default 0).

To evaluate reconstruction performance, run:

$ python evaluate_reconstruction.py --model-path $MODEL_PATH --config-path $CONFIG_PATH

To optimize the relative tolerance for the baseline in our paper, run:

$ python tune_tolerance.py --model-path $MODEL_PATH --config-path $CONFIG_PATH

We used 0.008 for KITTI.

Demo

$ streamlit run demo.py $MODEL_PATH $CONFIG_PATH

Synthesis	Reconstruction

Citation

If you find this code helpful, please cite our paper:

@inproceedings{nakashima2021learning,
    title     = {Learning to Drop Points for LiDAR Scan Synthesis},
    author    = {Nakashima, Kazuto and Kurazume, Ryo},
    booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
    pages     = {222--229},
    year      = 2021
}