⚡ electrical engineering

mRadNet: A Compact Radar Object Detector with MetaFormer

The paper introduces mRadNet, a compact and efficient radar object detection model leveraging a U-Net architecture with MetaFormer blocks to achieve state-of-the-art performance on the CRUW dataset with the fewest parameters and lowest FLOPs.

Original authors: Huaiyu Chen, Fahed Hassanat, Robert Laganiere, Martin Bouchard

Published 2026-05-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Huaiyu Chen, Fahed Hassanat, Robert Laganiere, Martin Bouchard

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a car in a heavy fog or a snowstorm. Your eyes (cameras) might struggle to see what's ahead, but your car's radar is like a pair of super-powered ears that can "hear" objects through the weather. The problem is, the raw data from these radar ears is a messy, confusing jumble of numbers. To make sense of it, we need a computer brain (an AI model) to clean it up and tell us, "There's a car 50 meters ahead," or "A cyclist is to the left."

The paper introduces a new computer brain called mRadNet. Here is how it works, explained simply:

The Problem: Too Big and Too Slow

Previous attempts to build these radar brains had a major flaw: they were like trying to carry a giant, heavy backpack (a huge computer model) while trying to run a marathon (real-time driving).

Old Models (CNNs): These were like a person looking at a picture through a tiny keyhole. They could see fine details right in front of them but couldn't understand the whole picture or how things were connected across the scene. To fix this, engineers made the models bigger and heavier, which slowed them down.
Newer Models (Transformers): These were like a person with a wide-angle lens who could see the whole room at once. However, they were computationally expensive (very heavy) and still struggled to see the tiny, local details needed for safety.

The Solution: mRadNet (The "Smart Compact" Brain)

The authors created mRadNet, a model designed specifically to be compact (small and light) but smart (accurate). They built it using a concept called MetaFormer.

Think of MetaFormer as a hybrid team:

The Local Specialist (Convolution): Imagine a detective who is great at examining a single fingerprint or a small clue. This part of the model looks at tiny, local details in the radar data.
The Global Strategist (Token Mixer): Imagine a general who looks at the whole battlefield map to understand how different parts connect. This part looks at the big picture and long-range connections.

mRadNet combines these two into a single, efficient unit. Instead of using a heavy, complex "self-attention" mechanism (which is like asking every person in a stadium to talk to every other person at once), mRadNet uses a simpler "token mixer." It's like having a streamlined communication system that gets the job done without the noise and cost.

How It Works: The U-Net Shape

The model is shaped like a U (called a U-Net architecture).

The Left Side (Encoder): Imagine squeezing a sponge. The model takes the large, detailed radar image and compresses it down, keeping only the most important information. It's like summarizing a long book into a few key bullet points.
The Right Side (Decoder): Imagine expanding that sponge back out. The model takes those key bullet points and expands them back into a full picture, but this time, it knows exactly where the objects are.
The Skip Connections: There are "bridges" connecting the left side to the right side. This ensures that when the model expands the image back out, it doesn't forget the tiny, fine details it saw at the beginning.

Special Tricks for Efficiency

To make sure this brain stays small and fast, the authors added two clever tricks:

Smart Packing (Token Embedding): Instead of looking at every single piece of radar data individually, the model groups them together efficiently, like packing a suitcase tightly so you can fit more in less space.
Smart Merging (Token Merging): As the model processes the data, it merges neighboring pieces of information together. Imagine taking a grid of 4 small photos and merging them into 1 larger, more detailed photo. This reduces the amount of work the computer has to do without losing any important information.

The Results

The team tested mRadNet on a massive dataset of real-world driving scenarios called CRUW.

Performance: It found objects (cars, pedestrians, cyclists) more accurately than any previous model.
Efficiency: It did this while being the smallest and lightest model tested. It required the least amount of computer power (FLOPs) and had the fewest "parameters" (the internal settings that make the AI smart).

In short: mRadNet is a new, lightweight radar detector that is small enough to fit on a car's computer chip but smart enough to see through bad weather better than the heavy, bulky models used before. It proves you don't need a giant backpack to run a marathon; you just need the right gear.

Technical Summary: mRadNet

Problem Statement

Frequency-Modulated Continuous Wave (FMCW) radars are increasingly vital for Advanced Driver Assistance Systems (ADAS) due to their robustness against adverse weather conditions like rain, fog, and snow. However, processing radar data for Object Detection (ROD) presents specific challenges for real-time embedded systems. While Convolutional Neural Networks (CNNs) effectively extract local features from radar Radio Frequency (RF) images, they struggle to capture global dependencies due to limited receptive fields, often necessitating large, computationally expensive models to achieve acceptable performance. Conversely, Transformer-based models excel at modeling global dependencies but suffer from high computational costs due to self-attention mechanisms and often fail to capture local features as effectively as CNNs without extensive architectural modifications. Furthermore, existing solutions often overlook the critical requirements for model compactness and efficiency in embedded automotive environments.

Methodology

The authors propose mRadNet, a novel, compact radar object detection model designed with efficiency in mind. The architecture integrates the strengths of Convolutional Neural Networks and MetaFormer architectures within a U-Net style framework.

1. Overall Architecture

mRadNet adopts a U-Net style encoder-decoder structure with skip connections to produce hierarchical representations.

Encoder: The input consists of complex RF images (range-azimuth) across multiple frames and chirps. The encoder processes these through a stem that extracts short-term temporal features, converting complex images to real-valued 3D images. It then utilizes a series of MetaFormer and Transformer blocks for hierarchical feature extraction.
Decoder: The decoder mirrors the encoder, applying Transformer and MetaFormer blocks to refine features. It employs skip connections to concatenate encoder tokens with decoder tokens, preserving low-level details.
Output: The final stage uses linear interpolation for smooth upsampling to generate confidence maps, followed by a convolutional layer to output class-specific confidence maps.

2. Key Architectural Components

MetaFormer Blocks with Token Mixers: Instead of relying solely on self-attention or standard convolutions, mRadNet employs MetaFormer blocks. These blocks replace the self-attention module with Token Mixers.
- Convolutional Token Mixers: The authors utilize Depthwise Separable Convolutions (SepConv) as token mixers. This decomposes standard convolution into depthwise and pointwise operations, effectively capturing local features while maintaining channel mixing capabilities without the redundancy of standard convolutions paired with linear layers.
- Attention Token Mixers: Attention mechanisms are also integrated to capture global dependencies, allowing the model to balance local and global feature extraction.
Token Embedding: A specialized TokenEmbed module processes the input 6D tensors (Batch $\times$ Time $\times$ Chirps $\times$ Height $\times$ Width $\times$ Channels). It uses a convolutional layer to merge chirp dimensions (mimicking FFT behavior with learnable parameters) and a 3D convolution to group spatiotemporal pixels into tokens, creating a location-aware token array.
Token Merging: To reduce computational cost and facilitate a deeper architecture, the model employs a TokenMerging module in the encoder. Unlike standard pooling, this module uses a Token Rearrange layer (similar to the inverse of PixelShuffle) to restructure $2 \times 2$ spatial neighborhoods into single tokens. This reduces spatial resolution by a factor of 4 while increasing token dimension, preserving information density before a linear layer reduces the dimension back to a manageable size. The temporal dimension remains untouched during this process.

Key Contributions

Novel Architecture: The proposal of mRadNet, a MetaFormer-based, U-Net style architecture specifically tailored for ROD tasks. It replaces traditional CNN backbones and heads with MetaFormer blocks containing convolutional layers, creating a cohesive design that captures both local and global features.
Efficient Token Strategies: The introduction of optimized token embedding and token merging strategies. These strategies facilitate the effectiveness of token mixers and enhance model compactness, enabling deeper architectures with fewer parameters.

Experimental Results

The model was validated on the CRUW dataset, a large-scale automotive radar object detection dataset.

Performance: mRadNet achieved state-of-the-art (SOTA) performance with an Overall Average Precision (AP) of 86.72% and Average Recall (AR) of 91.18%.
- It outperformed the previous SOTA model (E-RODNet) by 1.26% in AP and 1.99% in AR.
- It demonstrated significant improvements in car detection, achieving a 4%+ increase in AP over other methods, attributed to enhanced temporal modeling.
- While E-RODNet held a slight edge in AP for pedestrians and cyclists (likely due to more extensive convolutional operations for local features), mRadNet achieved superior AR for these classes.
Efficiency: mRadNet achieved the best performance with the fewest parameters and the lowest FLOPs among all compared methods (CNNs, Transformers, and MetaFormers).
- It demonstrated the shortest training time and lowest inference latency among Transformer and MetaFormer-based models.
- The paper notes that MetaFormer-based models generally tend to be smaller than CNN and Transformer counterparts while achieving better performance.

Significance and Claims

The paper claims that mRadNet successfully addresses the trade-off between detection accuracy and computational efficiency in radar object detection. By leveraging the MetaFormer architecture, the model establishes a solid performance lower bound that allows for aggressive design optimizations (reducing size and cost) without sacrificing accuracy. The integration of separable convolutions and attention token mixers within a U-Net framework proves effective for capturing the necessary local and global features from sparse and noisy radar data. The work demonstrates that MetaFormers can outperform both CNNs and Transformers in ROD tasks while maintaining a compact design suitable for real-time embedded systems.