chrischen1023 / dsrnet

Official implementation for "Single Image Reflection Separation via Component Synergy"

DSRNet: Single Image Reflection Separation via Component Synergy (ICCV 2023)

📖 [Arxiv] [Supp.]
Qiming Hu, Xiaojie Guo
College of Intelligence and Computing, Tianjin University

Network Architecture

Data Preparation

Training dataset

• 7,643 images from the Pascal VOC dataset, center-cropped as 224 x 224 slices to synthesize training pairs.
• 90 real-world training pairs provided by Zhang et al.

Testing dataset

• 45 real-world testing images from CEILNet dataset.
• 20 real testing pairs provided by Zhang et al.
• 454 real testing pairs from SIR^2 dataset, containing three subsets (i.e., Objects (200), Postcard (199), Wild (55)).

Download all in one by Google Drive or 百度云.

Usage

Training

python train_sirs.py --inet dsrnet_l --model dsrnet_model_sirs --dataset sirs_dataset --loss losses --name dsrnet_l --lambda_vgg 0.01 --lambda_rec 0.2 --if_align --seed 2018 --base_dir "[YOUR DATA DIR]"

Evaluation

python eval_sirs.py --inet dsrnet_s --model dsrnet_model_sirs --dataset sirs_dataset --name dsrnet_s_test --hyper --if_align --resume --weight_path "./weights/dsrnet_s_epoch14.pt" --base_dir "[YOUR_DATA_DIR]"

More commands can be found in scripts.sh.

Trained weights

Download the trained weights by Google Drive or 百度云 and drop them into the "weights" dir.

Visual comparison on real20 and SIR^2

Impressive Restoration Quality of Reflection Layers