Research map / Detection and segmentation / 3D detection and driving
RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RadarMOSEVE, a radar-only transformer that jointly estimates ego-velocity and segments moving objects from 4D radar point clouds, using new radar self- and cross-attention and a Doppler loss.
Problem
Joint moving object segmentation (MOS) and ego velocity estimation (EVE) from sparse, noisy millimeter-wave radar point clouds, where LiDAR-derived methods underperform and radial velocity is underused.
Approach
A radar transformer with object attention (ball-query sampling), scenario attention (interval and FPS-based wide-range sampling), and cross-attention across two frames, which uses radar radial Doppler velocity; EVE is trained with a Doppler loss and its output is used to compensate radial velocities before MOS.
Key results
- On the authors' dataset, the method reaches 70.2% average MOS IoU, about 9.4 points above 4DMOS.
- EVE MAE of 0.182 m/s and MSE of 0.065 versus 0.330 and 0.175 for Point Transformer; precision at <0.5 m/s is 94.3%, more than 17 points above PT.
- Ablation shows velocity compensation adds about 4.6 points mIoU, and raw Doppler velocity adds about 4.5 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add object, scenario, and cross-attention (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 65.7 -> 70.2 | Helped |
| Add cross-attention to the attention-free setup (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 65.7 -> 66.9 | Helped |
| Add object attention to cross-attention (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 66.9 -> 69.0 | Helped |
| Add scenario attention to cross-attention (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 66.9 -> 67.5 | Helped |
| Add scenario attention to object and cross-attention (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 69.0 -> 70.2 | Helped |
| Add cross-attention to object and scenario attention (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 66.6 -> 70.2 | Helped |
| Add raw radar velocity input (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 61.1 -> 65.6 | Helped |
| Add velocity compensation to raw velocity input (ablation) | RadarMOSEVE · Our dataset MOS | mIoU: 65.6 -> 70.2 | Helped |
| Use raw radar velocity (ablation) | RadarMOSEVE · Our dataset EVE | MAE: 0.301 -> 0.182 | Helped |
| Use raw radar velocity (ablation) | RadarMOSEVE · Our dataset EVE | MSE: 0.161 -> 0.065 | Helped |
| Use RadarMOSEVE instead of 4DMOS | Our dataset MOS · Our dataset MOS | Avg IoU: 60.8 -> 70.2 | Helped |
| Use RadarMOSEVE instead of PT | Our dataset EVE · Our dataset EVE | Precision <0.5m/s: 76.5 -> 94.3 | Helped |
Limitations
- Ill-defined when the scene contains only moving objects with sparse radar returns, especially when object motion resembles the robot's.
- Ego-velocity and MOS rely on radar-only input, which can be sparse and noisy.
- Evaluation on VoD uses only about 3,000 re-annotated frames after re-splitting.
Experimental setup
Datasets: View-of-Delft (VoD) Own radar MOSEVE dataset (13,654 frames)
Benchmarks: Own radar MOSEVE dataset View-of-Delft (VoD)
Metrics: IoU F1 accuracy MAE MSE precision at velocity thresholds
Baselines: ICP RANSAC Point Transformer Stratified-Transformer Point-BERT 4DMOS RaFlow CMFlow
Builds on: Point Transformer 4DMOS Stratified Transformer Point-BERT
Closest papers by meaning
- TransLoc4D: Transformer-based 4D Radar Place Recognition (CVPR 2024 · similarity 0.85)
- ICP-Flow: LiDAR Scene Flow Estimation with ICP (CVPR 2024 · similarity 0.84)
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.84)
- MITO: A Millimeter-Wave Dataset and Simulator for Non-Line-of-Sight Perception (NeurIPS 2026 · similarity 0.83)
- CRT-Fusion: Camera, Radar, Temporal Fusion Using Motion Information for 3D Object Detection (NeurIPS 2024 · similarity 0.83)
- RadarDistill: Boosting Radar-based Object Detection Performance via Knowledge Distillation from LiDAR Features (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.