Research map / Detection and segmentation / Open-set and few-shot detection
Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining
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.
Shows that detecting objects directly from RAW Bayer data (via a packed-input pipeline, PRD) with a new Functional Regularization activation improves detection accuracy and corruption robustness in low-light and harsh conditions, and introduces the RAW Corruption Benchmark (RCB).
Problem
Object detectors perform poorly in harsh environments (low light, weather, corruptions), and RGB images produced by the ISP discard RAW information useful for recognition; there is no realistic RAW corruption benchmark.
Approach
Skip the ISP and feed packed 4-channel RAW Bayer data (normalized over an expanded bit-depth range) into off-the-shelf detectors with only the first layer adapted, plus a bounded non-monotonic activation (Functional Regularization, combining ReLU6 and H-Swish) that exploits RAW's high bit depth.
Key results
- PRD with RAW-Packed 14-bit input improves CenterNet on nighttime LOD from 25.0 to 28.0 mAP over RGB, and EfficientNet-B3 from 24.5 to 29.1 mAP.
- On corrupted LOD (RCB), RAW outperforms RGB by 6.95 mAP for FCOS and 7.96 mAP for DETR on average; on LOD-Snow, PRD with FCOS is 13.9 mAP higher than RGB.
- CenterNet inference speed rises from 55.6 to 18.4 img/s... reported as 3.02x faster for RAW-Packed versus RGB (RAW 18.4 vs RGB 55.6 img/s as written).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PRD with RAW-Packed instead of RGB input | CenterNet · LOD night time | mAP: 25.0 to 28.0 | Helped |
| Use RAW rather than RGB input | FCOS · LOD Snow, severity=2 | mAP: 24.8 to 38.7 | Helped |
| Use RAW rather than RGB input | DETR · LOD Snow, severity=2 | mAP: 20.0 to 36.2 | Helped |
| Use RAW rather than RGB input | FCOS · LOD normal | mAP: 44.4 to 44.0 | No clear effect |
| Use RAW-Packed rather than RGB input (ablation) | CenterNet · LOD preprocessing ablation, range [0,1) | mAP: 25.0 to 23.7 | Hurt |
| Use RAW-Packed rather than RAW-Bayer input (ablation) | CenterNet · LOD preprocessing ablation, range [0,1) | mAP: 21.2 to 23.7 | Helped |
| Expand RAW-Packed normalization range from [0,1) to [0,64) (ablation) | CenterNet · LOD preprocessing ablation | mAP: 23.7 to 28.0 | Helped |
| Replace ReLU with FR (ablation) | CenterNet with RAW-Packed · LOD, range [0,4) | mAP: 25.7 to 28.5+2.8% | Helped |
| Replace ReLU with FR (ablation) | CenterNet with RAW-Packed · LOD, range [0,8) | mAP: 26.7 to 27.9+1.2% | Helped |
| Replace ReLU with FR (ablation) | CenterNet with RAW-Packed · LOD, range [0,64) | mAP: 28.0 to 28.7+0.7% | Helped |
| Replace ReLU with FR (ablation) | CenterNet with RAW-Bayer · LOD, range [0,1) | mAP: 21.2 to 21.9+0.7% | Helped |
| Replace ReLU with FR (ablation) | CenterNet with RAW-RGB · LOD, range [0,1) | mAP: 22.8 to 25.1+2.3% | Helped |
Surprising findings
- Packing RAW Bayer into 4 channels outperforms demosaiced RAW-RGB and standard RGB inputs, indicating demosaicing discards useful information.
- Authors report RAW inference at 18.4 img/s versus RGB at 55.6 img/s, which contradicts the claimed 3.02x speedup for RAW.
Limitations
- RGB inverse transformation could not be established, so unprocessed RAW vs RGB comparisons were not drawn.
- Pre-training with RAW-Packed is unfair since detectors are pretrained on 3-channel sRGB; RAW may need larger datasets for pre-training.
- Benchmark RCB is built from only three small datasets (LOD, RAW-NOD, PASCALRAW) and relies on synthetic corruptions.
Experimental setup
Datasets: LOD RAW-NOD PASCALRAW MS COCO
Benchmarks: LOD LOD-Snow RAW Corruption Benchmark (RCB) PASCALRAW RAW-NOD
Metrics: mAP mAP@50 mAP@75 composite corruption performance (CCP) inference speed (img/s)
Baselines: RGB detection RAW-Bayer input RAW-RGB input ReLU H-Swish ReLU6 DETR EfficientNet-B3 detector
Builds on: CenterNet FCOS Swish MMDetection
Closest papers by meaning
- RAWild: Toward Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling (NeurIPS 2026 · similarity 0.89)
- Towards RAW Object Detection in Diverse Conditions (CVPR 2025 · similarity 0.88)
- ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge (CVPR 2025 · similarity 0.86)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.84)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.82)
- Logarithmic Lenses: Exploring Log RGB Data for Image Classification (CVPR 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.