mRadNet: A Compact Radar Object Detector with MetaFormer

This is the official implementation of mRadNet from the paper mRadNet: A Compact Radar Object Detector with MetaFormer.

Data Preparation

• Download the CRUW ROD2021 dataset from https://www.cruwdataset.org/download. Download TRAIN_RAD_H.zip and TRAIN_RAD_H_ANNO.zip, the camera images and the testing set are not needed. Extract the zip files, and place the files as the following structure:

├─ annotations
|  ├─ test                      // 4 sequences - TRAIN_RAD_H_ANNO.zip
|  |  ├─ 2019_04_09_BMS1001.txt
|  |  └─ ...
|  └─ train                     // 36 sequences - TRAIN_RAD_H_ANNO.zip
|     ├─ 2019_04_09_BMS1000.txt
|     └─ ...
└─ sequences
   ├─ test                      // 4 sequences - TRAIN_RAD_H.zip
   |  ├─ 2019_04_09_BMS1001
   |  |  └─ RADAR_RA_H
   |  |     ├─ 000000_0000.npy
   |  |     └─ ...
   |  └─ ...
   └─ train                     // 36 sequences - TRAIN_RAD_H.zip
      |─ 2019_04_09_BMS1000
      |  └─ RADAR_RA_H
      |     ├─ 000000_0000.npy
      |     └─ ...
      └─ ...

Important

The testing sequences are 2019_04_09_BMS1001, 2019_04_30_MLMS001, 2019_05_23_PM1S013, 2019_09_29_ONRD005, same as T-RODNet and other models.

Installation

• Clone this repository
• Create a conda environment. mRadnet was tested under Python 3.12 on Ubuntu with Nvidia GPU.
• Install the required packages in requirements.txt.

  [!NOTE] For the cruw-devkit package, please refer to intructions here. It has not been maintained in recent years so modifications might be needed for it to work.

• Edit the paths in mRadNet.yaml to your own paths.

Training

python train.py -c mRadNet.yaml

Testing

python test.py -c mRadNet.yaml -r checkpoint.pt

Acknowledgement

• Thanks to Fahed Hassanat, Robert Laganière and Martin Bouchard for their instructions and resources.
• Thanks to MetaFormer authors for the lighweight archetecture.
• Thanks to the Yizhou Wang team for making part of the CRUW dataset public.
• Thanks to other research teams including T-RODNet authors and E-RODNet authors for their contributions to the ROD field.