Research map / Detection and segmentation / Open-set and few-shot detection
Exploiting Polarized Material Cues for Robust Car Detection
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 PCDNet, a YOLOv7-style two-branch RGB plus trichromatic polarization (AoLP/DoLP) detector with a new pixel-aligned RGBP-Car dataset for robust car detection.
Problem
Car detection in challenging conditions (low light, weather, dense traffic) where RGB color cues are unstable or limited, causing failures such as mirrored or black cars being missed or misdetected.
Approach
Fuse trichromatic linear polarization cues (AoLP and DoLP) with RGB through Polarization Integration, Material Perception modules that learn material features across samples, and a Cross-Domain Demand Query module that lets RGB request complementary polarization features in a request-and-complement manner.
Key results
- PCDNet reaches 58.5 AP, 85.2 AP50 and 61.5 AP75 on RGBP-Car, above YOLOv7 (57.6 AP) and YOLOv8-l (56.8 AP).
- RGB-only ablation drops to 57.6 AP, while adding DoLP to RGB gives 58.3 AP, indicating DoLP contributes more than AoLP.
- Removing the Material Perception module reduces AP from 58.5 to 56.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add trichromatic AoLP and DoLP to RGB input (ablation) | PCDNet · RGBP-Car test | AP: 57.6 -> 58.5 | Helped |
| Add DoLP rather than AoLP to RGB input (ablation) | PCDNet · RGBP-Car test | AP: 58.0 -> 58.3 | Helped |
| Stack AoLP and DoLP with RGB in the channel dimension (ablation) | PCDNet · RGBP-Car test | AP: 57.6 -> 57.5 | Hurt |
| Use RGB and trichromatic polarization instead of grayscale-only input (ablation) | PCDNet · RGBP-Car test | AP: 57.4 -> 58.5 | Helped |
| Enhance RGB with ZeroDCE instead of adding polarization (ablation) | PCDNet · RGBP-Car test | AP: 58.0 -> 57.4 | Hurt |
| Use PI to fuse AoLP and DoLP instead of simple concatenation (ablation) | PCDNet · RGBP-Car test | AP: 58.2 -> 58.5 | Helped |
| Use PI instead of adding or multiplying AoLP and DoLP (ablation) | PCDNet · RGBP-Car test | AP: 58.1 -> 58.5 | Helped |
| Add the Material Perception module (ablation) | PCDNet · RGBP-Car test | AP: 56.9 -> 58.5 | Helped |
| Apply MSP to shallow features and MCP to deeper features (ablation) | PCDNet · RGBP-Car test | AP: 58.2 -> 58.5 | Helped |
| Add the Cross-Domain Demand Query module (ablation) | PCDNet · RGBP-Car test | AP: 58.0 -> 58.5 | Helped |
| Add Spatial Demand Map Delivery to CDDQ (ablation) | PCDNet · RGBP-Car test | AP: 58.2 -> 58.5 | Helped |
| Add Channel Weight Dynamic Assignment to CDDQ (ablation) | PCDNet · RGBP-Car test | AP: 58.3 -> 58.5 | Helped |
Surprising findings
- Stacking AoLP and DoLP directly with RGB channels does not improve performance (57.5 AP vs 57.6 RGB-only).
- Enhancing low-light RGB with ZeroDCE (57.4 AP) is less effective than adding polarization cues.
Limitations
- Performance declines when both RGB and polarization signals are weak, e.g., headlight-induced holes at night.
- Limited ability to handle extreme HDR or heavy motion blur.
- Dataset is limited to 1,611 training images.
Experimental setup
Datasets: RGBP-Car ImageNet
Benchmarks: RGBP-Car
Metrics: AP AP50 AP75
Baselines: Faster R-CNN Cascade R-CNN Deformable DETR DINO YOLOX YOLOv7 YOLOv8 RTMDet P-SSD EfficientDet
Builds on: YOLOv7 CSPDarkNet Scharr edge extractor ZeroDCE
Closest papers by meaning
- RGB-to-Polarization Estimation: A New Task and Benchmark Study (NeurIPS 2025 · similarity 0.85)
- Glass Segmentation Using Intensity and Spectral Polarization Cues (CVPR 2022 · similarity 0.85)
- Pre-training LiDAR-based 3D Object Detectors through Colorization (ICLR 2024 · similarity 0.83)
- DALDet: Depth-Aware Learning Based Object Detection for Autonomous Driving (AAAI 2024 · similarity 0.82)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.82)
- A New Multi-Source Light Detection Benchmark and Semi-Supervised Focal Light Detection (NeurIPS 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.