Recasting Regional Lighting for Shadow Removal

Authors: Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, Rynson W.H. Lau

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three benchmarks show that our method outperforms existing SOTA shadow removal methods.
Researcher Affiliation Academia Yuhao Liu, Zhanghan Ke, Ke Xu*, Fang Liu, Zhenwei Wang, Rynson W.H. Lau Department of Computer Science, City University of Hong Kong {yuhao Liu7456, kkangwing, fawnliu2333}@gmail.com, {zhanghake2-c, zhenwwang2-c}@my.cityu.edu.hk, rynson.lau@cityu.edu.hk
Pseudocode No The paper describes algorithmic steps but does not include formal pseudocode blocks or algorithm listings.
Open Source Code No The paper does not provide a link to open-source code or explicitly state its release.
Open Datasets Yes We conduct experiments on three shadow removal datasets, i.e., SRD (Qu et al. 2017), ISTD (Wang, Li, and Yang 2018), and ISTD+ (Le and Samaras 2021).
Dataset Splits No SRD consists of 3,088 paired shadow and shadow-free images, which are split into 2680 for training and 408 for testing. ISTD contains 1,870 shadow images, shadow masks, and shadow-free image triplets, of which 1,330 are used for training and 540 for testing. (No explicit mention of validation splits for any dataset).
Hardware Specification Yes Our method is implemented via the Py Torch Toolbox on a single NVIDIA TESLA V100 GPU with 32G memory
Software Dependencies No The paper mentions 'Py Torch Toolbox' but does not specify a version number or other software dependencies with versions.
Experiment Setup Yes The initial learning rate, β1, β2, and batch size being set to 0.0002, 0.9, 0.999, and 12. Learning rate adjustment utilizes a warmup and cosine decay strategy. For the local diffusion process, we set the times steps T, initial and end variance scheduler βt to {1000,0.0001,0.02} and {50,0.0001,0.5} for training and testing. Data is augmented by random flipping and cropping, and resized to 256 256 for training. Shadow-aware decomposition and bilateral correction networks are trained for 100k and 200k iterations, respectively.