Research map / Detection and segmentation / CNN features
TARSS-Net: Temporal-Aware Radar Semantic Segmentation Network
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 TARSS-Net, a radar semantic segmentation network with a Temporal Relation-Aware Module (target-history relation encoder TH-TRE plus learnable temporal relation attentive pooling TRAP) that aggregates radar temporal information for multi-view segmentation.
Problem
Radar semantic segmentation (RSS) under-exploits temporal information in radar sequences; existing methods rely on 3D convolution or simple pooling, and Transformers are too costly along time for high-dimensional radar tensors.
Approach
Encode relations between the target frame and each historical frame (TH-TRE, O(T) with shared kernels) and aggregate them with a learnable attentive pooling (TRAP, in spatio-temporal or depth-temporal form) instead of 3DConv or GAP/GMP, keeping parameter count constant across variable input lengths.
Key results
- On CARRADA, TARSS-Net_D reaches 63.4% RD mIoU and 75.2% RD mDice, versus 60.7% / 72.6% for PKCIn-Net, with 6.3M parameters.
- On CARRADA RA view, TARSS-Net_D (41.4 mIoU) trails PKCIn-Net (43.1) and TransRSS (43.0), which the authors attribute to weak temporal cues in RA data.
- On CARRADA-RAC global view, TARSS-Net_D reaches 55.1% mIoU versus 54.0% for PKCIn-Net; on KuRALS (single-view RD), TARSS-Net_D reaches 58.4% mIoU versus 56.7% for PKCIn-Net_sv.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TARSS-Net_D instead of PKCIn-Net | CARRADA-RAC models · CARRADA-RAC Global | mIoU: 54.0% -> 55.1% | Helped |
| Use TARSS-Net_D instead of PKCIn-Netsv | KuRALS models · KuRALS RD view | mIoU: 56.7% -> 58.4% | Helped |
| Add TRAM instead of using stacked 3D convolutions (ablation) | TARSS-Net · CARRADA test, RD view | mIoU: 56.1 -> 63.4 | Helped |
| Add TH-TRE to the GAP baseline (ablation) | Baseline-A w/ GAP · CARRADA test, RD view | mIoU: 55.4 -> 60.6 | Helped |
| Add TH-TRE rather than replacing it with regular 3D convolution (ablation) | TARSS-Net · CARRADA test, RD view | mIoU: 61.6 -> 63.4 | Helped |
| Use Depth-TRAP instead of GAP (ablation) | TARSS-Net · CARRADA test, Global view | mIoU: 50.5 -> 52.4 | Helped |
| Use Spatio-TRAP instead of GAP (ablation) | TARSS-Net · CARRADA test, Global view | mIoU: 50.5 -> 51.9 | Helped |
| Use Depth-TRAP instead of Spatio-TRAP (ablation) | TARSS-Net · CARRADA test, Global view | mIoU: 51.9 -> 52.4 | Helped |
| Add Depth-TRAP to Baseline-A (ablation) | Baseline-A · CARRADA test, Global view | mIoU: 45.9 -> 51.0 | Helped |
| Add Spatio-TRAP to Baseline-A (ablation) | Baseline-A · CARRADA test, Global view | mIoU: 45.9 -> 49.9 | Helped |
| Add AD encoding branch (ablation) | Baseline-B · CARRADA test, RD view | mIoU: 52.8 -> 56.1 | Helped |
| Add AD encoding branch (ablation) | TARSS-Net_D · CARRADA test, RD view | mIoU: 60.1 -> 63.4 | Helped |
Surprising findings
- A ViT-based temporal baseline performed worse than the 3DConv-based TMVA-Net while being far more expensive (27M params, 12 FPS multi-view).
- On the RA view, the proposed TARSS-Net is not the best; the authors attribute this to weak temporal information in RA data.
Limitations
- Trails PKCIn-Net and TransRSS on the RA view, where temporal information is less obvious.
- Inference speed (23 FPS multi-view) leaves room for improvement.
- The KuRALS dataset is not yet fully released; the paper says the complete dataset is forthcoming.
Experimental setup
Datasets: CARRADA CARRADA-RAC KuRALS
Benchmarks: CARRADA CARRADA-RAC KuRALS
Metrics: mIoU mDice precision recall FPS MACs parameters
Baselines: FCN U-Net DeepLabv3+ RSS-Net RAMP-CNN MV-Net MVA-Net TMVA-Net TransRadar T-RODNet TransRSS PKCIn-Net
Builds on: TMVA-Net PKCIn-Net 3D convolution self-attention
Closest papers by meaning
- AdaPKC: PeakConv with Adaptive Peak Receptive Field for Radar Semantic Segmentation (NeurIPS 2024 · similarity 0.82)
- TASeg: Temporal Aggregation Network for LiDAR Semantic Segmentation (CVPR 2024 · similarity 0.81)
- SIRA: Scalable Inter-frame Relation and Association for Radar Perception (CVPR 2024 · similarity 0.80)
- RadarMOSEVE: A Spatial-Temporal Transformer Network for Radar-Only Moving Object Segmentation and Ego-Velocity Estimation (AAAI 2024 · similarity 0.80)
- RADAR: Relative Angular Divergence Across Representations (NeurIPS 2026 · similarity 0.79)
- FM-ChangeNet: Learning Change through Pathwise Feature Transport (NeurIPS 2026 · similarity 0.79)
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.