Research map / Detection and segmentation / Tracking
SIRA: Scalable Inter-frame Relation and Association for Radar Perception
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 SIRA, a radar-only detection and tracking framework combining extended temporal relation over many frames (window-based cross-frame attention) with motion-consistent pseudo-tracklet association, evaluated on Radiate.
Problem
Radar perception for oriented object detection and multi-object tracking suffers from low resolution, noise, ghost targets and nonlinear motion, and prior methods use only one or two frames.
Approach
Generalizes the two-frame temporal relation layer to multi-frame temporal windows with Swin-style temporally regrouped window attention for scalable cross-frame attention, and adds a learned pseudo-direction head whose pseudo-tracklets drive a KF-based association similarity.
Key results
- SIRA-34 (4 frames) reaches 58.11 mAP@0.5 on Radiate (train good weather), +4.11 over prior state of the art.
- SIRA-34 with OC-SORT reaches 47.79 MOTA, +9.94 over prior best reported in the table comparisons.
- Using four frames instead of two gives roughly +5 mAP@0.3 and +4 mAP@0.5 over TempoRadar two-frame detection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SIRA's extended temporal relation and motion-consistency association | ResNet-34 radar perception · Radiate, train good weather | mAP@0.5: 57.01±1.03 -> 58.11±1.40 | Helped |
| Add SIRA's extended temporal relation and motion-consistency association | ResNet-34 radar perception · Radiate, train good & bad weather | mAP@0.5: 51.99±1.94 -> 53.79±1.14 | Helped |
| Add pseudo-direction estimation loss during training | ETR-34 (4) · Radiate, train good weather | MOTA: 47.11 -> 47.79 | Helped |
| Use max rather than mean or sum patch merging (ablation) | SIRA-34 (4) · Radiate, detection ablation | mAP@0.5: 55.06±2.07; 55.55±1.59; 56.47±1.54 | Helped |
| Increase masked MCA counts from H1=1, H2=1 to H1=2, H2=1 (ablation) | SIRA-34 (4) · Radiate, detection ablation | mAP@0.5: 56.65±2.38 -> 57.59±0.84 | Helped |
| Increase TRWA masked MCA count from H2=1 to H2=2 (H1=1) (ablation) | SIRA-34 (4) · Radiate, detection ablation | mAP@0.5: 56.65±2.38 -> 58.46±0.91 | Helped |
| Increase ETR stages from L=1 to L=4 (ablation) | SIRA-34 (4) · Radiate, detection ablation | mAP@0.3: 68.68±1.12 -> 69.16±1.06 | Helped |
| Add Ctracklet alone to association (ablation) | SIRA-34 (4) + OC-SORT · Radiate, train good weather | MOTA: 47.11 -> 47.11 | No clear effect |
| Add Cangle alone to association (ablation) | SIRA-34 (4) + OC-SORT · Radiate, train good weather | MOTA: 47.11 -> 47.00 | Hurt |
| Use both Ctracklet and Cangle for association (ablation) | SIRA-34 (4) + OC-SORT · Radiate, train good weather | MOTA: 47.11 -> 47.79 | Helped |
Limitations
- Evaluated only on a single radar dataset (Radiate)
- Relies on high-density radar detection points and a KF-based tracker
- Center crop of 256x256 used for main results, full-size results only in appendix
Experimental setup
Datasets: Radiate
Benchmarks: Radiate
Metrics: mAP@0.3 mAP@0.5 MOTA IDF1 IDs Frag MT PT
Baselines: RetinaNet CenterPoint BBAVectors TempoRadar SeTR CenterTrack OC-SORT
Builds on: TempoRadar Swin Transformer CenterPoint OC-SORT CenterTrack
Closest papers by meaning
- RETR: Multi-View Radar Detection Transformer for Indoor Perception (NeurIPS 2024 · similarity 0.84)
- Temporal Coherent Object Flow for Multi-Object Tracking (AAAI 2025 · similarity 0.83)
- GRAE-3DMOT: Geometry Relation-Aware Encoder for Online 3D Multi-Object Tracking (CVPR 2025 · similarity 0.82)
- HGSFusion: Radar-Camera Fusion with Hybrid Generation and Synchronization for 3D Object Detection (AAAI 2025 · similarity 0.82)
- Modeling Continuous Motion for 3D Point Cloud Object Tracking (AAAI 2024 · similarity 0.82)
- RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation (AAAI 2024 · similarity 0.82)
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.