Paper Critique Reading List¶

Why Learning on Spectrum?¶

RAD Tensor Detection ICCVW 2019

Vehicle Detection With Automotive Radar Using Deep Learning on Range-Azimuth-Doppler Tensors

Qualcomm's early detector that runs on the dense range-azimuth-Doppler tensor instead of the peak-detected point cloud.
TMVA-Net ICCV 2021

Multi-View Radar Semantic Segmentation

Segments the range-angle-Doppler tensor through several of its 2D views, beating image-derived and radar baselines on CARRADA with far fewer parameters.
GRADE MobiCom 2026

GRADE: Single-Frame Generative Radar Depth Estimation Under Visual Degradation

Conditions a pretrained latent diffusion model on coarse depth from raw 4D radar spectra to produce metric depth that holds up in real smoke.

FFT-RadNet CVPR 2022

Raw High-Definition Radar for Multi-Task Learning

Learns angles directly from HD-radar range-Doppler spectra for vehicle detection and free-space segmentation, and introduces the RADIal dataset.
ADCNet arXiv 2023

ADCNet: Learning from Raw Radar Data via Distillation

A learnable signal-processing layer, initialized as a DFT, runs directly on raw ADC data ahead of a backbone, pre-trained by distilling the classical radar pipeline.
SSMRadNet WACV 2026

SSMRadNet: A Sample-wise State-Space Framework for Efficient and Ultra-Light Radar Segmentation and Object Detection

Two stacked Mamba SSMs read raw ADC sample by sample (fast time) and then chirp by chirp (slow time), matching FFT-RadNet-level segmentation on RADIal with about 60x less compute.

Echoes Beyond Points NeurIPS 2023

Echoes Beyond Points: Unleashing the Power of Raw Radar Data in Multi-modality Fusion

EchoFusion skips the point-cloud pipeline and fuses raw radar spectrum features with camera features through BEV queries.
CRN ICCV 2023

CRN: Camera Radar Net for Accurate, Robust, Efficient 3D Perception

Uses sparse radar points to lift camera features into BEV, then fuses the two with multi-modal deformable attention; runs at 20 FPS and rivals LiDAR detectors on nuScenes.
TacoDepth CVPR 2025 Oral

TacoDepth: Towards Efficient Radar-Camera Depth Estimation with One-stage Fusion

One-stage radar-camera depth estimation that fuses graph features of sparse radar points into an image pyramid, skipping intermediate quasi-dense depth.

Soli SIGGRAPH 2016

Soli: Ubiquitous Gesture Sensing with Millimeter Wave Radar

Google's end-to-end 60 GHz radar sensor and pipeline for sub-millimeter fine gesture tracking.
Wi-Vi SIGCOMM 2013

See Through Walls with WiFi!

Uses MIMO interference nulling to cancel static reflections and treats a moving body as an antenna array, detecting people and gestures behind walls with Wi-Fi-band USRP radios.
mmMesh MobiSys 2021

mmMesh: Towards 3D Real-Time Dynamic Human Mesh Construction Using Millimeter-Wave

Real-time 3D human mesh estimation from sparse point clouds of a commercial single-chip mmWave radar, using a built-in body model to reach about 2.5 cm vertex error.

MIMO-SAR IEEE TVT 2021

MIMO-SAR: A Hierarchical High-Resolution Imaging Algorithm for mmWave FMCW Radar in Autonomous Driving

Applies coherent SAR to a vehicle-mounted TI AWR1843 MIMO radar for sharper side-view azimuth resolution, using a two-stage hierarchical pipeline and radar odometry to estimate the trajectory.
PanoRadar MobiCom 2024 Best Demo Award

Enabling Visual Recognition at Radio Frequency

A rotating single-chip mmWave radar on a robot forms a dense synthetic aperture, with learning-based elevation enhancement, to produce LiDAR-like 3D images for segmentation and detection.
X-band NLOS CVPR 2026

X-band Radar Non-Line-of-Sight Imaging

Uses X-band's long wavelength to make wall bounces mostly specular, then reconstructs hidden geometry with a learned dense predictor plus geometry-aware NLOS module, reaching 40 m in experiments.

NeRF2 MobiCom 2023 Best Paper Award Runner-Up

NeRF2: Neural Radio-Frequency Radiance Fields

A radiance field for RF propagation that predicts the signal received anywhere from a known transmitter, then generates synthetic training data for downstream networks.
DART CVPR 2024 Oral

DART: Implicit Doppler Tomography for Radar Novel View Synthesis

A NeRF-style implicit scene model with a radar-physics renderer that synthesizes range-Doppler images from novel views.
Radar Fields SIGGRAPH 2024

Radar Fields: Frequency-Space Neural Scene Representations for FMCW Radar

Neural occupancy and reflectance fields learned in Fourier frequency space from raw FMCW radar, recovering dense 3D geometry from a single drive.