RAWild

Official implementation of RAWild: Sensor-Agnostic RAW Object Detection via Physics-Guided Curve and Grid Modeling.

RAWild adds a lightweight RAW adapter before the downstream vision backbone. The adapter predicts a per-image Bezier tone curve and a bilateral-grid color transform, enabling robust perception across RAW images from different sensors, exposure levels, bit depths, and spectral responses.

This release includes object detection, semantic segmentation, RAW simulation, and adapter visualization code.

Assets

Datasets and checkpoints will be released separately.

• Dataset download:
  • RAWild_Mutiraw: Google Drive
  • Simulation: Google Drive

Dataset folders include RAWild_Mutiraw and Simulation under the Drive dataset folder.

• Checkpoint download: Google Drive

Checkpoint folders are organized as Checkpoints/Det for object detection and Checkpoints/Seg for semantic segmentation. Detection checkpoint files are named as <Backbone>-<Dataset>.pth, for example ResNet50-PAS.NM.pth; segmentation checkpoint files are named as <Backbone>-<Domain>.pth, for example MiT-B3-normal.pth.

Set paths before running the examples:

export RAWILD_DATA_ROOT=<RAWild_datasets>
export RAWILD_PASCAL_NPY_ROOT="$RAWILD_DATA_ROOT/PASCAL_RAW_npy"
export RAWILD_PRETRAINED_ROOT=<checkpoints/pretrained>
export RAWILD_RELEASE_CKPT_ROOT=<checkpoints/released>
export RAWILD_WORK_DIR_ROOT=work_dirs

export RAWILD_MMSEG_DATA_ROOT="$RAWILD_DATA_ROOT/ADEChallengeData2016"
export RAWILD_MMSEG_WORK_DIR_ROOT=work_dirs/rawild_mmseg

RAWILD_DATA_ROOT should contain PASCAL_RAW, PASCAL_RAW_npy, LOD_BMVC2021, ROD_dataset, AODRaw_dataset, and ADEChallengeData2016.

Installation

Detection:

conda create -n RAWild_mmdet python=3.8 -y
conda activate RAWild_mmdet

cd mmdetection_github
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
pip install -r requirements.txt
pip install -v -e .

Segmentation:

conda create -n RAWild_mmseg python=3.8 -y
conda activate RAWild_mmseg

cd mmsegmentation_github
pip install -r requirements.txt
pip install -e . --no-deps

The segmentation dependency snapshot was verified with Python 3.8, PyTorch 2.1.0 + CUDA 12.1, MMCV 2.1.0, and MMEngine 0.10.4.

Object Detection

Configs:

Setting	ResNet-50	Swin-T
PASCALRAW normal	config/RAWild_resnet50/pas_nm.py	config/RAWild_swint/pas_nm.py
PASCALRAW low-light	config/RAWild_resnet50/pas_low.py	config/RAWild_swint/pas_low.py
PASCALRAW over-exposure	config/RAWild_resnet50/pas_oe.py	config/RAWild_swint/pas_oe.py
LOD	config/RAWild_resnet50/lod.py	config/RAWild_swint/lod.py
PASCALRAW + LOD mix	config/RAWild_resnet50/pascal_lod_mix.py	config/RAWild_swint/pascal_lod_mix.py
ROD	config/RAWild_resnet50/rod.py	config/RAWild_swint/rod.py
AODRaw	config/RAWild_resnet50/aodraw.py	config/RAWild_swint/aodraw.py

Reported results of RAWild:

Backbone	PAS.LOW @50/@75	PAS.NM @50/@75	PAS.OE @50/@75	LOD @50/@75	ROD @50/@75	AODRAW @50/@75
ResNet-50	0.8911 / 0.7332	0.9019 / 0.7690	0.9020 / 0.7690	0.6909 / 0.4363	0.5541 / 0.3752	0.3623 / 0.2311
Swin-T	0.9083 / 0.7406	0.9342 / 0.7792	0.9313 / 0.7921	0.6986 / 0.5057	0.5191 / 0.3467	0.4394 / 0.3263

Train:

cd mmdetection_github

python tools/train.py config/RAWild_resnet50/pascal_lod_mix.py
python tools/train.py config/RAWild_swint/pascal_lod_mix.py

Distributed training:

bash tools/dist_train.sh config/RAWild_resnet50/pascal_lod_mix.py 4
bash tools/dist_train.sh config/RAWild_swint/pascal_lod_mix.py 4

Evaluate:

python tools/test.py \
  config/RAWild_resnet50/pas_nm.py \
  "$RAWILD_RELEASE_CKPT_ROOT/ResNet50-PAS.NM.pth"

python tools/test.py \
  config/RAWild_swint/pas_nm.py \
  "$RAWILD_RELEASE_CKPT_ROOT/SwinT-PAS.NM.pth"

Use the config table above to switch datasets or backbones.

Semantic Segmentation

Configs:

Backbone	Normal	Low	Over Exposure	Mix
MiT-B0	config/RAWild_mitb0/normal.py	config/RAWild_mitb0/low.py	config/RAWild_mitb0/oe.py	config/RAWild_mitb0/mix.py
MiT-B3	config/RAWild_mitb3/normal.py	config/RAWild_mitb3/low.py	config/RAWild_mitb3/oe.py	config/RAWild_mitb3/mix.py
MiT-B5	config/RAWild_mitb5/normal.py	config/RAWild_mitb5/low.py	config/RAWild_mitb5/oe.py	config/RAWild_mitb5/mix.py

Reported results of RAWild:

Backbone	LOW mIoU	NM mIoU	OE mIoU
MiT-B0	0.2872	0.3534	0.3372
MiT-B3	0.3957	0.4516	0.4381
MiT-B5	0.4082	0.4708	0.4560

Train:

cd mmsegmentation_github

python tools/train.py config/RAWild_mitb0/mix.py
python tools/train.py config/RAWild_mitb3/mix.py
python tools/train.py config/RAWild_mitb5/mix.py

Distributed training:

bash tools/dist_train.sh config/RAWild_mitb0/mix.py 4
bash tools/dist_train.sh config/RAWild_mitb3/mix.py 4
bash tools/dist_train.sh config/RAWild_mitb5/mix.py 4

Evaluate:

python tools/test.py \
  config/RAWild_mitb0/mix.py \
  "$RAWILD_RELEASE_CKPT_ROOT/rawild_mitb0_mix.pth"

python tools/test.py \
  config/RAWild_mitb3/mix.py \
  "$RAWILD_RELEASE_CKPT_ROOT/rawild_mitb3_mix.pth"

python tools/test.py \
  config/RAWild_mitb5/mix.py \
  "$RAWILD_RELEASE_CKPT_ROOT/rawild_mitb5_mix.pth"

For paper-style evaluation, use the single-scale sliding-window path in the config and do not enable TTA.

RAW Simulation

The released dataset already includes processed PASCALRAW .npy files under PASCAL_RAW_npy. If you start from original PASCAL RAW .nef files, preprocess them with:

cd mmdetection_github

python PASCAL_RAW_pre_process.py \
  --raw-root <PASCAL_RAW_NEF_DIR> \
  --out-root "$RAWILD_DATA_ROOT/PASCAL_RAW/original"

Generate synthetic spectrum and mixed-bit PASCALRAW data:

python tools/syn_spec/generate_pascal_normal_mixbit.py \
  --source-root "$RAWILD_PASCAL_NPY_ROOT/demosaic_normal_12bit" \
  --train-list "$RAWILD_DATA_ROOT/PASCAL_RAW/trainval/train.txt" \
  --val-list "$RAWILD_DATA_ROOT/PASCAL_RAW/trainval/val.txt" \
  --output-root "$RAWILD_WORK_DIR_ROOT/syn_spec/pascal_normal_mixbit" \
  --bit-depths 8 9 10 11 12

Visualize camera-seed variants:

python tools/syn_spec/visualize_camera_seed_views.py \
  --sample-id 2014_000022 \
  --source-root "$RAWILD_PASCAL_NPY_ROOT/demosaic_normal_12bit" \
  --output-root "$RAWILD_WORK_DIR_ROOT/syn_spec/camera_seed_views"

Visualization

Input	Bezier

Bezier + Grid	Curve

Generate adapter visualizations:

cd mmdetection_github

NO_ALBUMENTATIONS_UPDATE=1 python tools/visual_selection/visualize_bezier.py \
  --config config/RAWild_resnet50/pas_nm.py \
  --checkpoint "$RAWILD_RELEASE_CKPT_ROOT/ResNet50-PAS.NM.pth" \
  --image-key 2014_000001 \
  --out-dir "$RAWILD_WORK_DIR_ROOT/visual_selection/pas_nm"

Export the Bezier curve only:

NO_ALBUMENTATIONS_UPDATE=1 python tools/visual_selection/visualize_bezier.py \
  --config config/RAWild_resnet50/pas_nm.py \
  --checkpoint "$RAWILD_RELEASE_CKPT_ROOT/ResNet50-PAS.NM.pth" \
  --image-key 2014_000001 \
  --out-dir "$RAWILD_WORK_DIR_ROOT/visual_selection/pas_nm_curve" \
  --curve-only

Citation

If you find this work useful, please cite:

@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}
}