Research map / Detection and segmentation / CNN features

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

NeurIPS 2026 · Detection and segmentation › CNN features · approach: Context for hard detection 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.

Proposes RAWild, a histogram-conditioned global Bézier tone curve plus spatially adaptive bilateral grid adapter for sensor-agnostic RAW object detection, with a physics-based RAW simulation pipeline for augmentation.

Problem

RAW object detection degrades across sensors because exposure, spectral sensitivity, and bit depth vary, creating large domain gaps; existing task-oriented ISPs are tied to specific sensor configurations.

Approach

A histogram-guided shared transformer predicts a delta-residual Bézier tone curve (global, per-channel monotone) and a bilateral grid of structured D(I+A) color matrices (local), placed as a near-identity adapter before a pretrained detector backbone; a physics-based RAW simulator generates sensor-varied training data.

Key results

• On ResNet-50, Ours reaches 0.6909 mAP@50 on LOD and 0.5541 on ROD, versus the strongest baseline's 0.6408 and 0.4720 (Dr.RAW / AdaptiveISP).
• On Swin-Transformer, Ours reaches 0.9313 mAP@50 on PASCALRAW overexposed and 0.4394 on AODRAW.
• Ours reports 9.8% relative mAP@50 gain over the strongest baseline on Multi-RAW mixed-sensor data (0.3528 vs 0.2572 for Dr.RAW).

Surprising findings

• A vanilla absolute-control-point Bézier curve falls slightly below the no-curve baseline on the LOD ablation (0.6511 vs 0.6650 mAP@50).

Limitations

• Real mixed-sensor RAW data with balanced labels is scarce, limiting large-scale experiments.
• MultiRAW required custom rebalancing to a 5-class subset, so results are not directly comparable to the original benchmark.

Experimental setup

Datasets: PASCALRAW LOD ROD AODRAW MultiRAW ADE20K

Benchmarks: PASCALRAW LOW/NM/OE LOD ROD AODRAW PASCALRAW (Syn) LOD (Syn) ROD (Syn) PAS & LOD Multi-RAW ADE20K (RAW segmentation)

Metrics: mAP@50 mAP@75 mIoU

Baselines: Linear-RAW Default-ISP ReconfigISP RAW-Adapter AdaptiveISP Dark-ISP Dr.RAW

Builds on: Bilateral Grid HDRnet RAW-Adapter Dr.RAW RetinaNet

Closest papers by meaning

• Guiding a Harsh-Environments Robust Detector via RAW Data Characteristic Mining (AAAI 2024 · similarity 0.89)
• ReRAW: RGB-to-RAW Image Reconstruction via Stratified Sampling for Efficient Object Detection on the Edge (CVPR 2025 · similarity 0.86)
• Towards RAW Object Detection in Diverse Conditions (CVPR 2025 · similarity 0.86)
• Cubify Anything: Scaling Indoor 3D Object Detection (CVPR 2025 · similarity 0.84)
• PhysAug: A Physical-guided and Frequency-based Data Augmentation for Single-Domain Generalized Object Detection (AAAI 2025 · similarity 0.83)
• Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network (AAAI 2025 · similarity 0.83)