ILLRAW: Illumination-Aware RAW Object Detection

Xinzhe Wang; Xiao Liang; Xiaoli Zhang; Kang Ma

cs.CVNov 2, 2026

ILLRAW: Illumination-Aware RAW Object Detection

Authors: Xinzhe Wang, Xiao Liang, Xiaoli Zhang, Kang Ma

Organizations: Beijing Institute of Technology · Independent Researcher · State Key Laboratory of CNS/ATM, Beijing Institute of Technology

Abstract

Existing RAW-based detectors typically introduce learnable ISP modules to convert RAW inputs into sRGB-like formats to fit off-the-shelf detectors. However, this RAW-to-sRGB paradigm decouples representation learning from the downstream detection task. Crucially, it forces the network to inherit human-centric visual priors while discarding intrinsic RAW advantages, such as linear sensor responses and high-bit-depth, minimally processed scene cues. In this paper, we propose ILLRAW, an ILLumination-aware end-to-end RAW object detection framework. Instead of reconstructing visually plausible sRGB images, ILLRAW directly learns detection-oriented RAW features through a lightweight feature adapter inserted at the input stage of existing detectors. ILLRAW estimates an illumination mask and introduces channel-sensitive embeddings to model Bayer-channel responses. These features are fused by an illumination attention fusion module. Experiments show clear improvements on the diverse AODRaw benchmark and competitive performance on LOD and PascalRAW, with particularly strong results under dark exposure.

Figures & tables

Panel of Figure 1 — Figure 1: Comparison of sRGB and RAW images. ( a ) RAW pixel values are concentrated within a small range. ( b ) RAW images exhibit significant channel differences. ( c ) In the example scene, the green channel is brighter, while the red channel is darker.

Figure 2: Overview of the proposed ILLRAW . It leverages illumination variations across the channel and spatial dimensions to effectively extract semantic information from RAW images.

Detector	Method	mAP	AP 50	AP 75	Normal	Low	Rain	Fog	Params	FPS
C-RCNN	sRGB	18.4	29.5	19.5	21.0	16.1	15.7	14.4	60.96	40.8
C-RCNN	RAW*	18.1	30.9	18.3	20.6	16.7	17.3	14.6	60.96	40.8
C-RCNN	Karaimer et al.	18.3	29.7	19.6	22.1	16.1	16.1	13.8	60.96	40.8
C-RCNN	InvISP	17.4	28.7	18.3	21.3	15.6	13.1	13.0	61.10	30.7
C-RCNN	RAOD	21.1	33.6	22.2	23.8	19.3	19.9	17.4	61.03	35.6
C-RCNN	RAW-Adapter*	19.9	33.2	20.1	22.3	18.1	20.8	16.9	61.11	36.2

Table 1: Comparison results on the AODRaw dataset. We use mAP and AP under different weather conditions to evaluate the detection performance. Results marked with * correspond to the values reported in [ 27 ] . Here, C-RCNN and D-DETR denote Cascade-RCNN and Deformable DETR. Params and FPS denote the number of model parameters in millions and the inference speed measured on one NVIDIA RTX 3090 GPU, respectively.

Method	Bicycle	Car	Motor.	Chair	Dining	Bottle	TV	Bus	Total
sRGB*	67.6	89.1	59.9	77.3	39.8	34.8	50.9	55.9	59.4
RAW*	71.9	89.3	57.7	74.6	37.1	31.0	43.6	62.4	58.5
Karaimer et al. *	66.9	87.3	57.6	71.3	30.2	29.7	35.8	56.0	54.4
InvISP*	72.6	89.2	59.9	76.8	38.9	31.9	31.4	56.6	57.1
RAOD	71.0	89.9	28.2	83.4	32.4	31.2	34.8	63.4	54.2
RAW-Adapter*	76.2	89.7	58.0	77.4	40.6	43.2	46.0	65.9	62.1

Table 2: Comparison results on the LOD dataset. The detector is RetinaNet [ 29 ] with a ResNet-50 backbone. We report the AP for each category and the overall mAP. Results marked with * correspond to the values reported in [ 27 ] .

Method	Normal	Overexposure	Dark
sRGB*	88.3	-	-
RAW*	87.7	87.7	80.3
Karaimer et al. *	88.1	85.6	78.8
InvISP*	85.4	86.6	70.9
RAOD	88.1	88.1	83.9
RAW-Adapter*	88.7	88.7	82.5

Table 3: Comparison results on the PascalRAW dataset. The detector is RetinaNet [ 29 ] with a ResNet-18 backbone. We present AP under three different illumination conditions. Results marked with * correspond to the values reported in [ 7 ] .

Panel of Figure 3 — Figure 3: Visual examples of detection results with all the comparison methods. Please zoom in for confidence scores and class predictions.

Panel of Figure 4 — Figure 4: Visualization of the illumination mask $\mathbf{M}$ and its complement $1-\mathbf{M}$ . They effectively distinguish different brightness regions in the image, guiding the model to apply distinct operations for areas with varying illumination.

CSE		Fusion module				mAP
G	RB	Mask	Sum	Cat	IAF	mAP
-	-	-	-	-	-	16.8
✓	-	-	-	-	-	20.4
-	✓	-	-	-	-	21.1
✓	✓	-	✓	-	-	22.0
✓	✓	✓	✓	-	-	22.1

Table 4: Ablation study on AODRaw without gamma correction.

Method	mAP	AP 50	AP 75
Random	21.7	34.7	22.8
ResNet-18	22.4	36.1	24.0
sRGB	23.0	36.2	24.7
Synthesized RAW	23.3	36.0	24.8

Table 5: Impact of backbone pretraining on our method.

Low-Light Object Detection

See all 9 papers

RAWild: Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling
May 7, 2026Shuhong Liu, Gengjia Chang, Jun Liu +4Domain Generalization Object Detection
The University of Tokyo · I2WM · 12-bit 1/4000 360 ISO +1
FreqAdapt: Frequency-Adaptive Processing for RAW Object Detection
Aug 4, 2026Hanxi Li, Huiling LiImage Enhancement Frequency-Domain Feature Learning
University of Science and Technology China · Li Auto Inc · Hunan University
SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras
Mar 16, 2026Huanjing Yue, Shangbin Xie, Cong Cao +4Image-to-Image Translation Image Restoration
School of Electrical and Information Engineering, Tianjin University · State key laboratory of Smart Power Distribution Equipment and System, Tianjin University · Individual

May 7, 2026J/K move · Enter open · S save

RAWild: Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling

Shuhong Liu, Gengjia Chang, Jun Liu, Xuangeng Chu+3

The University of Tokyo · I2WM · 12-bit 1/4000 360 ISO+1

Domain Generalization Object Detection

Open paper AskPDF arXiv

Camera sensor RAW data offers intrinsic advantages for object detection, including deeper bit depth, preserved physical information, and freedom from image signal processor (ISP) distortions. However, varying exposure conditions, spectral sensitivities, and bit depths across devices introduce substantially larger domain gaps than sRGB, making sensor-agnostic generalization a fundamental challenge. In this study, we present RAWild, a physics-guided global-local tone mapping framework for sensor-agnostic RAW object detection. By factoring sensor-induced variations into a global tonal correction and a spatially adaptive local color adjustment, both driven by RAW distribution priors, our framework enables a single network to train jointly across heterogeneous sensors. To further support cross-sensor generalization, we construct a physics-based RAW simulation pipeline that synthesizes realistic sensor outputs spanning diverse spectral sensitivities, illuminants, and sensor non-idealities. Extensive experiments across multiple RAW benchmarks covering bit depths from 10 to 24 demonstrate state-of-the-art (SOTA) performance under single-dataset, mixed-dataset, and challenging robustness settings.

Abstract

Low-Light Object Detection

Explore similar work

RAWild: Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling

FreqAdapt: Frequency-Adaptive Processing for RAW Object Detection

SpiralDiff: Spiral Diffusion with LoRA for RGB-to-RAW Conversion Across Cameras