Research map / Detection and segmentation / CNN features
AdaPKC: PeakConv with Adaptive Peak Receptive Field for Radar Semantic Segmentation
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 AdaPKC, a PeakConv variant whose peak receptive field (guard bandwidth) adapts per cell via a metric-based or learned estimator, plus a thresholded fine-tuning switch (FiTOS), improving radar semantic segmentation on CARRADA and a self-collected Ku-band dataset.
Problem
Fixed peak receptive field in PeakConv cannot handle varying target broadening and non-homogeneous, time-varying clutter in radar semantic segmentation.
Approach
Replace PeakConv's fixed guard bandwidth with a per-cell adaptive PRF: AdaPKCξ selects the guard bandwidth by a feature-space correlation metric (max first-order gradient of sorted scores), while AdaPKCθ learns a continuous guard bandwidth via a small network with linear interpolation; FiTOS fine-tunes with a confidence threshold to switch adaptation on or off.
Key results
- On CARRADA RA view, AdaPKCξ-Net reaches 44.1% mIoU vs 43.1% for PKCIn-Net; AdaPKCξ-Net-FiT reaches 62.1% mIoU / 74.0% mDice in RD view vs 60.7% / 72.6% for PKCIn-Net.
- On KuRALS, KuRALS-Net w/ AdaPKCξ-FiT reaches 58.2% mIoU vs 56.7% for KuRALS-Net w/ PKC.
- FiTOS adds gains without extra training resources; AdaPKCξ-Net-FiT is best overall on CARRADA among the reported models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace fixed PeakConv receptive fields with metric-based adaptive receptive fields | PeakConv RSS model · CARRADA RA view | mIoU: 43.1% -> 44.1% | Helped |
| Replace fixed PeakConv receptive fields with learned adaptive receptive fields | PeakConv RSS model · CARRADA RD view | mIoU: 60.7% -> 61.5% | Helped |
| Add FiTOS fine-tuning to AdaPKCξ-Net | AdaPKCξ-Net · CARRADA RD view | mIoU: 61.2% -> 62.1% | Helped |
| Add AdaPKCξ with FiTOS | PKCIn-Net · CARRADA-RAC RA view | mIoU: 47.3% -> 48.8% | Helped |
| Add AdaPKC to KuRALS-Net | KuRALS-Net · KuRALS | mIoU: 56.0% -> 58.2% | Helped |
| Set PeakConv range guard bandwidth to 2 instead of 1 (ablation) | PeakConv · CARRADA RD view | mIoU: 60.7% -> 61.0% | Helped |
| Set PeakConv range guard bandwidth to 4 instead of 1 (ablation) | PeakConv · CARRADA RD view | mIoU: 60.7% -> 59.7% | Hurt |
| Use three-round Vot-MRT to enlarge the PRF search space (ablation) | AdaPKCξ-Net · CARRADA RD view | mIoU: 61.2% -> 61.8% | Helped |
| Use three-round Vot-MRT to enlarge the PRF search space (ablation) | AdaPKCξ-Net · CARRADA RA view | mIoU: 44.1% -> 44.6% | Helped |
| Increase the FiTOS threshold from its best setting (ablation) | AdaPKCξ-Net FiT · CARRADA RD and RA views | mDice | Hurt |
Surprising findings
- On CARRADA-RAC, AdaPKCθ-Net has lower RD-view mIoU (60.7%) than AdaPKCξ-Net (61.6%), and the gains over PKCIn-Net are modest in several settings.
- Manually widening the guard bandwidth in PeakConv does not monotonically help; bandwidth 2 is best and larger values degrade performance.
Limitations
- Adaptive PRF search is restricted to a local range due to GPU memory limits (motivating the Vot-MRT extension).
- Gains over PKC are moderate on several metrics, and results are reported on only two radar datasets.
- The FiTOS threshold tau must be tuned, with performance declining when it is too large or too small.
Experimental setup
Datasets: CARRADA CARRADA-RAC KuRALS CRUW
Benchmarks: CARRADA CARRADA-RAC KuRALS
Metrics: mIoU mDice FPS GMACs
Baselines: PeakConv PKCIn-Net TMVA-Net TransRadar T-RODNet RAMP-CNN RSS-Net FCN U-Net DeepLabv3+ DefConv DefConvV2
Builds on: PeakConv PKCIn-Net CA-CFAR deformable convolution TMVA-Net
Closest papers by meaning
- TARSS-Net: Temporal-Aware Radar Semantic Segmentation Network (NeurIPS 2024 · similarity 0.82)
- RadSimReal: Bridging the Gap Between Synthetic and Real Data in Radar Object Detection With Simulation (CVPR 2024 · similarity 0.81)
- Bootstrapping Autonomous Driving Radars with Self-Supervised Learning (CVPR 2024 · similarity 0.81)
- SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar Object Detection (AAAI 2025 · similarity 0.81)
- RETR: Multi-View Radar Detection Transformer for Indoor Perception (NeurIPS 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)
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.