4DMOS

Our moving object segmentation on the unseen SemanticKITTI test sequences 18 and 21. Red points are predicted as moving.

Please find the corresponding video here.

Given a sequence of point clouds, our method segments moving (red) from non-moving (black) points.

We first create a sparse 4D point cloud of all points in a given receding window. We use sparse 4D convolutions from the MinkowskiEngine to extract spatio-temporal features and predict per-points moving object scores.

Important Update

Please also check out our latest map-based moving object segmentation pipeline MapMOS.

The current state of the repository is improved by internally aligning the scans using KISS-ICP. Also, the build system and pipeline have been updated after the release of our newer method MapMOS, so you can run it on most point cloud data formats. If you want to reproduce the original results from the paper, this version is tagged under 0.1. You can checkout by

git checkout v0.1

Installation

First, make sure the MinkowskiEngine is installed on your system, see here for more details.

Next, clone our repository

git clone [email protected]:PRBonn/4DMOS && cd 4DMOS

and install with

make install

or

make install-all

if you want to install the project with all optional dependencies (needed for the visualizer). In case you want to edit the Python code, install in editable mode:

make editable

How to Use It

Just type

mos4d_pipeline --help

to see how to run 4DMOS.

Check the Download section for a pre-trained model. Like KISS-ICP, our pipeline runs on a variety of point cloud data formats like bin, pcd, ply, xyz, rosbags, and more. To visualize these, just type

mos4d_pipeline --visualize /path/to/weights.ckpt /path/to/data

Because these labels come in all shapes, you need to specify a dataloader. This is currently available for SemanticKITTI, NuScenes, HeLiMOS, and our labeled KITTI Tracking sequence 19 and Apollo sequences (see Downloads).

Training

To train our approach, you need to first cache your data. To see how to do that, just cd into the 4DMOS repository and type

python3 scripts/precache.py --help

After this, you can run the training script. Again, --help shows you how:

python3 scripts/train.py --help

You can inspect the cached training samples by using the script python3 scripts/cache_to_ply.py --help.

The training log and checkpoints will be saved by default to the current working directory. To change that, export the export LOGS=/your/path/to/logs environment variable before running the training script.

HeLiMOS

To train on the HeLiMOS data with different sensor configurations, use the following commands:

python3 scripts/precache.py /path/to/HeLiMOS helimos /path/to/cache --config config/helimos/*_training.yaml
python3 scripts/train.py /path/to/HeLiMOS helimos /path/to/cache --config config/helimos/*_training.yaml

by replacing the paths and the config file names. To evaluate for example on the Velodyne test data, run

mos4d_pipeline /path/to/weights.ckpt /path/to/HeLiMOS --dataloader helimos -s Velodyne/test.txt

Evaluation and Visualization

We use the SemanticKITTI API to evaluate the intersection-over-union (IOU) of the moving class as well as to visualize the predictions. Clone the repository in your workspace, install the dependencies and then run the following command to visualize your predictions for e.g. sequence 8:

cd semantic-kitti-api
./visualize_mos.py --sequence 8 --dataset /path/to/dataset --predictions /path/to/4DMOS/predictions/ID/POSES/labels/STRATEGY/

Benchmark

To submit the results to the LiDAR-MOS benchmark, please follow the instructions here.

Downloads

• Model [A]: 5 scans @ 0.1s
• Model [B]: 5 scans @ 0.2s
• Model [C]: 5 scans @ 0.3s
• Model [D]: 5 scans, no poses
• Model [E]: 5 scans input, 1 scan output
• Model [F]: 2 scans
• Model [G]: 10 scans

Publication

If you use our code in your academic work, please cite the corresponding paper:

@article{mersch2022ral,
author = {B. Mersch and X. Chen and I. Vizzo and L. Nunes and J. Behley and C. Stachniss},
title = {{Receding Moving Object Segmentation in 3D LiDAR Data Using Sparse 4D Convolutions}},
journal={IEEE Robotics and Automation Letters (RA-L)},
year = 2022,
volume = {7},
number = {3},
pages = {7503--7510},
codeurl = {https://github.com/PRBonn/4DMOS},
}

Acknowledgments

This implementation is heavily inspired by KISS-ICP.

License

This project is free software made available under the MIT License. For details see the LICENSE file.