RAW Object Detection

Installation

Please install the conda environment as described in README. Afterward, mmdetection can be installed by following the official installation steps: https://mmdetection.readthedocs.io/en/3.x/get_started.html.

cd object_detection
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0,<2.2.0"

pip install -v -e .

Experiments

Object detection experiments are performed on NOD Nikon and Sony.

NOD Nikon

RGB

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_rgb_pIN.py --cfg-options randomness.seed=0

RAW

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=1.0 randomness.seed=0

RAW + Cityscapes-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=0.05 randomness.seed=0

RAW + BDD100K-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_raw3c_pIN_bddraw.py --cfg-options train_dataloader.dataset.f=0.05 randomness.seed=0

NOD Sony

RGB

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_rgb_pIN.py --cfg-options randomness.seed=0

RAW

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=1.0 randomness.seed=0

RAW + Cityscapes-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=0.05 randomness.seed=0

RAW + BDD100K-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_raw3c_pIN_bddraw.py --cfg-options train_dataloader.dataset.f=0.05 randomness.seed=0

Zero-Shot Experiments

NOD Nikon

Cityscapes-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=0.0 randomness.seed=0

BDD100K-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_nikon_raw3c_pIN_bddraw.py --cfg-options train_dataloader.dataset.f=0.0 randomness.seed=0

NOD Sony

Cityscapes-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_raw3c_pIN_csraw.py --cfg-options train_dataloader.dataset.f=0.0 randomness.seed=0

BDD100K-RAW (RAW-Diffusion)

python tools/train.py raw/configs/faster-rcnn_r50_fpn_nod_h416_sony_raw3c_pIN_bddraw.py --cfg-options train_dataloader.dataset.f=0.0 randomness.seed=0