PASCAL RAW dataset results

Object Detection on PASCAL RAW dataset.

PASCAL RAW: Download PASCAL RAW dataset from [Google Drive] or [百度网盘 (passwd: oiab)]

Which format as:

--  data
     -- PASCAL_RAW_github
         -- annotations
         -- original (original RAW, demosaic RAW normal-light & over-exposure & low-light)
         -- compare_ISP (ISP methods, InvISP, ECCV16-ISP)
         -- trainval

LOD dataset Results

Object Detection on LOD dataset.

LOD: Download LOD dataset from Google Drive or 百度网盘 (passwd: kf43), or find the LOD dataset original provide link.

Which format as:

--  data
     -- LOD_BMVC21
         -- RAW_dark (RAW data, "demosacing" in our paper)
         -- RGB_dark (default ISP RGB data)
         -- RAW_dark_InverseISP (InvISP processed RAW data, [CVPR 2021])
         -- RAW_dark_ECCV16_Micheal (ECCV16 ISP processed RAW data, [ECCV 2016])
         -- RAW-dark-Annotations (detection label)      
         -- trainval

ADE20K-RAW dataset results

Our proposed ADE 20K RAW dataset download link:
Google Drive or 百度网盘 (passwd: acv7)

Dataset Format:

--  data
     -- ADE20K
         -- ADEChallengeData2016
             -- annotations
                 -- training
                 -- validation
             -- images

                 # Synthesis RAW data
                 -- training_raw (synthesis RAW)
                 -- validation_raw
                 -- training_raw_low  (synthesis RAW, low-light)
                 -- validation_raw_low
                 -- training_raw_over_exp  (synthesis RAW, over-exposure)
                 -- validation_raw_over_exp

                 # Compare ISP methods
                  -- ISP_methods
                      -- 1_SID (https://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_Learning_to_See_CVPR_2018_paper.pdf)
                      -- 2_InvISP (https://yzxing87.github.io/InvISP/index.html)
                      -- 3_ECCV16 (https://karaimer.github.io/camera-pipeline/)

We release the ADE 20K RAW segmentation weights and training logs:

1. RAW-Adapters (SegFormer with MIT-B5/B3/B0 backbone) :
   normal-light (normal) & over-exposure (oe) & low-light (low)

2. Demosaic (SegFormer with MIT-B5 backbone) :
   normal-light (normal) & over-exposure (oe) & low-light (low)

3. Other Compare ISP Methods (SegFormer with MIT-B5 backbone):

InvISP (CVPR 2021): normal-light & over-exposure & low-light
ECCV 16 ISP (ECCV 2016): normal-light & over-exposure & low-light
SID (CVPR 1028): low-light