RAWild
arXiv Code

NeurIPS 2026

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

Shuhong Liu1,2* Gengjia Chang2* Jun Liu2 Xuangeng Chu1,2

Yinqiang Zheng1 Tatsuya Harada1,3 Ziteng Cui1†

1The University of Tokyo 2I2WM 3RIKEN

* Equal contribution † Corresponding author

Paper Code Dataset BibTeX
The film
0:00 / 1:12

The filmRAWild in seventy-two seconds.

01The problem

Every sensor sees the world differently.

Exposure, spectral response and bit depth change from camera to camera, and a detector trained on one camera's RAW stumbles on the next.

PASCALRAW scene as recorded by a Nikon D3200
12bitNikon D3200
AODRaw scene as recorded by a Sony α7 III
14bitSony α7 III
LOD scene as recorded by a Canon 5D IV
16bitCanon 5D IV
ROD scene as recorded by a Sony IMX490
24bitSony IMX490

02Overview

The full architecture.

RAWild architecture: a RAW image and its quantile histogram drive two transformer blocks via AdaLN, jointly predicting a global Bézier curve and a bilateral grid that feed the backbone network.
Figure 4 A RAW image and its quantile histogram drive transformer blocks via AdaLN, jointly predicting a global Bézier curve and a bilateral grid for local color correction.

03The method

A global curve, then a local grid.

A light adapter in front of the detector, steered by the RAW histogram.

  1. 1Bézier curve

    One tone curve per channel, pinned to black and white.

  2. 2Bilateral grid

    A 3×3 color matrix for every cell and brightness level.

  3. 3Histogram guidance

    A 64-quantile descriptor steers both, sensor by sensor.

The adapter's input: a linear RAW night frame, almost black The same frame after the Bézier curve The same frame after the bilateral grid, as the detector sees it
  1. Linear RAWWhat the sensor records. At night, almost black.
  2. Bézier curveOne curve per channel lifts the tones.
  3. Bilateral gridColor is finished locally, cell by cell.

ROD · night-07732 · stage outputs of the adapter

04Results

What the baselines miss.

The same frame for every method, one score threshold.

foundmissedfalse alarm

Mixed-sensor detection

One network, every sensor.

Multi-RAW

35.28mAP@50

+9.25over the strongest baseline

LOD
64.0869.09
ROD
47.2055.41
RAW ADE20K segmentation
9 / 9

mAP@50 in percent. Multi-RAW: Table 2. LOD and ROD: strongest baseline (Dr.RAW) → RAWild, ResNet-50, Table 1. Segmentation: best mIoU in all nine settings, Table 7.

All numbers from the paperTables 1, 2, 3, 7

First page of the RAWild paper

05The paper

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

NeurIPS 2026

@article{liu2026rawild,
  title     = {RAWild: Sensor-Agnostic RAW Object
               Detection via Physics-Guided Curve and
               Grid Modeling},
  author    = {Liu, Shuhong and Chang, Gengjia and
               Liu, Jun and Chu, Xuangeng and
               Zheng, Yinqiang and Harada, Tatsuya and
               Cui, Ziteng},
  journal   = {arXiv preprint arXiv:2605.05941},
  year      = {2026}
}