Awesome
Multi-View Radar Semantic Segmentation
Paper
Multi-View Radar Semantic Segmentation, ICCV 2021.
Arthur Ouaknine, Alasdair Newson, Patrick Pérez, Florence Tupin, Julien Rebut
This repository groups the implemetations of the MV-Net and TMVA-Net architectures proposed in the paper of Ouaknine et al..
The models are trained and tested on the CARRADA dataset.
The CARRADA dataset is available on Arthur Ouaknine's personal web page at this link: https://arthurouaknine.github.io/codeanddata/carrada.
If you find this code useful for your research, please cite our paper:
@InProceedings{Ouaknine_2021_ICCV,
author = {Ouaknine, Arthur and Newson, Alasdair and P\'erez, Patrick and Tupin, Florence and Rebut, Julien},
title = {Multi-View Radar Semantic Segmentation},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2021},
pages = {15671-15680}
}
Installation with Docker
It is strongly recommanded to use Docker with the provided Dockerfile containing all the dependencies.
- Clone the repo:
$ git clone https://github.com/ArthurOuaknine/MVRSS.git
- Create the Docker image:
$ cd MVRSS/
$ docker build . -t "mvrss:Dockerfile"
Note: The CARRADA dataset used for train and test is considered as already downloaded by default. If it is not the case, you can uncomment the corresponding command lines in the Dockerfile or follow the guidelines of the dedicated repository.
- Run a container and join an interactive session. Note that the option
-v /host_path:/local_path
is used to mount a volume (corresponding to a shared memory space) between the host machine and the Docker container and to avoid copying data (logs and datasets). You will be able to run the code on this session:
$ docker run -d --ipc=host -it -v /host_machine_path/datasets:/home/datasets_local -v /host_machine_path/logs:/home/logs --name mvrss --gpus all mvrss:Dockerfile sleep infinity
$ docker exec -it mvrss bash
Installation without Docker
You can either use Docker with the provided Dockerfile containing all the dependencies, or follow these steps.
- Clone the repo:
$ git clone https://github.com/ArthurOuaknine/MVRSS.git
- Install this repository using pip:
$ cd MVRSS/
$ pip install -e .
With this, you can edit the MVRSS code on the fly and import function and classes of MVRSS in other project as well.
-
Install all the dependencies using pip and conda, please take a look at the Dockerfile for the list and versions of the dependencies.
-
Optional. To uninstall this package, run:
$ pip uninstall MVRSS
You can take a look at the Dockerfile if you are uncertain about steps to install this project.
Running the code
In any case, it is mandatory to specify beforehand both the path where the CARRADA dataset is located and the path to store the logs and models. Example: I put the Carrada folder in /home/datasets_local, the path I should specify is /home/datasets_local. The same way if I store my logs in /home/logs. Please run the following command lines while adapting the paths to your settings:
$ cd MVRSS/mvrss/utils/
$ python set_paths.py --carrada /home/datasets_local --logs /home/logs
Training
In order to train a model, a JSON configuration file should be set. The configuration file corresponding to the selected parameters to train the TMVA-Net architecture is provided here: MVRSS/mvrss/config_files/tmvanet.json
. To train the TMVA-Net architecture, please run the following command lines:
$ cd MVRSS/mvrss/
$ python train.py --cfg config_files/tmvanet.json
If you want to train the MV-Net architecture (baseline), please use the corresponding configuration file: mvnet.json
.
Testing
To test a recorded model, you should specify the path to the configuration file recorded in your log folder during training. Per example, if you want to test a model and your log path has been set to /home/logs
, you should specify the following path: /home/logs/carrada/tmvanet/name_of_the_model/config.json
. This way, you should execute the following command lines:
$ cd MVRSS/mvrss/
$ python test.py --cfg /home/logs/carrada/tmvanet/name_of_the_model/config.json
Note: the current implementation of this script will generate qualitative results in your log folder. You can disable this behavior by setting get_quali=False
in the parameters of the predict()
method of the Tester()
class.
Acknowledgements
- The paper is under review, special thanks will be indicated after the final results
- The Soft Dice loss is based on the code from https://github.com/kornia/kornia/blob/master/kornia/losses/dice.py
License
The MVRSS repo is released under the Apache 2.0 license.