Research map / Detection and segmentation / Open-set and few-shot detection
Towards RAW Object Detection in Diverse Conditions
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.
Introduces AODRaw, a 7,785-image real RAW object detection dataset across 9 light/weather conditions with 62 categories, and proposes RAW pre-training on synthesized ImageNet-RAW with cross-domain distillation from an sRGB teacher, avoiding ISP adapters.
Problem
Object detection is usually done on sRGB images compressed by an ISP, losing information useful for detection in adverse light and weather; existing RAW datasets are small and low in category/condition diversity, and sRGB pre-training transfers poorly to RAW.
Approach
Pre-train the backbone directly on RAW images (synthesized ImageNet-RAW via unprocessing, with gamma correction) and use logit and feature distillation from an sRGB-pretrained teacher to overcome the difficulty of RAW pre-training, removing the need for trainable neural ISP adapters.
Key results
- RAW pre-training with distillation lifts ConvNeXt-T + Cascade RCNN from 34.0% AP (sRGB) to 34.8% AP on AODRaw, with gains in adverse conditions (+4.8% APrain).
- sRGB pre-trained Cascade RCNN fine-tuned on RAW drops to 33.7% AP, showing the sRGB-RAW domain gap.
- Distillation on ImageNet-RAW reaches 81.8% Top-1 for ConvNeXt-T versus 82.1% for sRGB pre-training, and degrades less under brightness and noise shifts.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RAW-domain pre-training with cross-domain distillation instead of sRGB pre-training | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | AP: 34.0 -> 34.8 | Helped |
| Use RAW-domain pre-training with cross-domain distillation instead of sRGB pre-training | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | APrain: 32.9 -> 36.1 | Helped |
| Use RAW-domain pre-training instead of sRGB pre-training | Cascade R-CNN + ConvNeXt-T · AODRaw test (sliced) | AP: 29.9 -> 30.7 | Helped |
| Add cross-domain distillation (ablation) | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | AP: 34.1 -> 34.8 | Helped |
| Add logit distillation (ablation) | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | AP: 34.1 -> 34.3 | Helped |
| Add feature distillation alongside logit distillation (ablation) | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | AP: 34.3 -> 34.8 | Helped |
| Use RAW pre-training instead of sRGB pre-training | Cascade R-CNN + ConvNeXt-T · AODRaw test (down-sampled) | AP: 33.7 -> 34.8 | Helped |
| Add a trainable pre-processing module | YOLOv8-n · AODRaw test (down-sampled, real-time detection) | FPS: 188.1 -> 57.6 | Hurt |
Surprising findings
- sRGB pre-training transfers worse to RAW than RAW-domain pre-training, so sRGB pre-training constrains RAW detection performance.
- Adding synthetic noise to ImageNet-RAW lowers pre-training accuracy (74.8% to 74.4%), indicating noise harms RAW pre-training.
Limitations
- Synthetic ImageNet-RAW uses unprocessing rather than real RAW pre-training data, and pre-training accuracy remains below sRGB pre-training.
- Gains are modest (0.8% AP over sRGB pre-training in the main setting).
- Dataset is captured with a single camera (Sony A7M4), limiting sensor diversity.
Experimental setup
Datasets: AODRaw ImageNet-1K COCO RAOD
Benchmarks: AODRaw
Metrics: AP AP50 AP75 APs APm APl APnormal APlow APrain APfog
Baselines: Faster RCNN Sparse RCNN RetinaNet GFocal Deformable DETR RAOD RAW-Adapter YOLOX-Tiny YOLOv6-n YOLOv8-n YOLO-MS-XS
Builds on: Cascade R-CNN ConvNeXt unprocessing (Brooks et al.) RAOD neural ISP RAW-Adapter
Closest papers by meaning
- Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining (AAAI 2024 · similarity 0.88)
- Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.86)
- RAWild: Toward Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling (NeurIPS 2026 · similarity 0.86)
- ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge (CVPR 2025 · similarity 0.85)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- Semi-supervised Open-World Object Detection (AAAI 2024 · similarity 0.85)
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.