Research map / Point clouds / Point cloud classification
TransLoc4D: Transformer-based 4D Radar Place Recognition
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 TransLoc4D, an end-to-end 4D radar place recognition network combining a sparse-convolution MinkLoc4D backbone (using intensity and ego-motion-derived velocity attributes) with a linear Transformer, and builds two 4D radar place recognition benchmarks.
Problem
Place recognition for autonomous vehicles in adverse weather using 4D millimeter-wave radar, which is sparse and noisy, with no prior open-source datasets or methods for 4D radar place recognition.
Approach
Refines radar scans via RANSAC ego-velocity regression to remove dynamic and noise points, converts Doppler velocity into a radar-motion-invariant azimuth attribute, feeds geometry+intensity+velocity into a sparse-conv backbone, and adds a linear self-attention Transformer after interpolation to capture long-range context before GeM pooling.
Key results
- TransLoc4D-R-VI-T reaches Recall@1 of 86.8% on NYL-Rain and 90.8% on SJTU4DPR-TestA, versus 81.0% and 88.6% for MinkLoc3Dv2 trained on NTU4DPR.
- Compared with MinkLoc3Dv2, TransLoc4D improves by more than 6% on NYL-Rain and 5% on SRC-Night in Recall@1 according to the paper's text.
- Ablation shows the refinement, velocity-azimuth attribute, intensity, and Transformer each add gains, with the full model about 10% above the plain MinkLoc4D baseline on benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add point cloud refinement (ablation) | TransLoc4D · NYL-Night, NYL-Rain, SRC-Night, and SJTU4DPR-TestA | Recall@1: 93.2 -> 96.6; 75.8 -> 81.0; 86.2 -> 89.0; 84.6 -> 88.6 | Helped |
| add velocity azimuth feature (ablation) | TransLoc4D-R · NYL-Rain, SRC-Night, and SJTU4DPR-TestA | Recall@1: 81.0 -> 83.3; 89.0 -> 93.6; 88.6 -> 89.5 | Helped |
| add velocity azimuth feature (ablation) | TransLoc4D-R · NYL-Night | Recall@1: 96.6 -> 95.7 | Hurt |
| add intensity feature (ablation) | TransLoc4D-R-V · NYL-Night and SRC-Night | Recall@1: 95.7 -> 96.8; 93.6 -> 94.4 | Helped |
| add intensity feature (ablation) | TransLoc4D-R-V · NYL-Rain and SJTU4DPR-TestA | Recall@1: 83.3 -> 82.5; 89.5 -> 89.0 | Hurt |
| add Transformer feature enhancement (ablation) | TransLoc4D-R-VI · NYL-Night, NYL-Rain, SRC-Night, and SJTU4DPR-TestA | Recall@1: 96.8 -> 97.1; 82.5 -> 86.8; 94.4 -> 94.5; 89.0 -> 90.8 | Helped |
| use TransLoc4D instead of PTC-Net-L | 4D radar place recognition · NYL-Night | Recall@1: 96.6 -> 97.1 | Helped |
| use TransLoc4D instead of PTC-Net-L | 4D radar place recognition · NYL-Rain | Recall@1: 87.6 -> 86.8 | Hurt |
| use TransLoc4D instead of PTC-Net-L | 4D radar place recognition · SRC-Night | Recall@1: 94.5 -> 94.5 | No clear effect |
| use TransLoc4D instead of MinkLoc3Dv2 | 4D radar place recognition · SJTU4DPR-TestA | Recall@1: 88.6 -> 90.8 | Helped |
| use TransLoc4D instead of MinkLoc3Dv2 | 4D radar place recognition · SJTU4DPR-TestB | Recall@1: 85.8 -> 85.9 | No clear effect |
Surprising findings
- Pre-trained PTC-Net-L (on Oxford LiDAR) is sometimes inferior to MinkLoc3Dv2, suggesting LiDAR-specific learned patterns do not transfer to 4D radar.
Limitations
- Intensity is susceptible to rainfall and cross-domain differences, reducing gains on rainy and SJTU subsets in some ablations.
- Evaluated only on two radar datasets with a limited number of loops (three NTU trajectories, four SJTU).
Experimental setup
Datasets: NTU4DRadLM SJTU4D NTU4DPR SJTU4DPR
Benchmarks: NTU4DPR-NYL-Night NTU4DPR-NYL-Rain NTU4DPR-SRC-Night SJTU4DPR-TestA SJTU4DPR-TestB
Metrics: Recall@1 Recall@5 Recall@10
Baselines: Intensity Scan Context AutoPlace MinkLoc3Dv2 TransLoc3D PTC-Net-L
Builds on: MinkLoc3Dv2 Linear Transformer PTC-Net TSAP loss Intensity Scan Context
Closest papers by meaning
- Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.87)
- L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.87)
- RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar (NeurIPS 2024 · similarity 0.86)
- RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation (AAAI 2024 · similarity 0.85)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.84)
- SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection (AAAI 2025 · similarity 0.84)
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.