Efficient Model-Driven Network for Shadow Removal
Authors: Yurui Zhu, Zeyu Xiao, Yanchi Fang, Xueyang Fu, Zhiwei Xiong, Zheng-Jun Zha3635-3643
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our method could achieve SOTA performance with less than half parameters, one-fifth of floating-point of operations (FLOPs), and over seventeen times faster than SOTA method (DHAN). Extensive experiments indicate that our method achieves leading shadow removal performance in terms of quantitative metrics, inference efficiency, and visual quality. Ablation Study |
| Researcher Affiliation | Academia | 1 University of Science and Technology of China, China 2 University of Toronto, Canada |
| Pseudocode | Yes | Algorithm 1: Iterative algorithm for shadow removal. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing its source code or a link to a code repository. |
| Open Datasets | Yes | ISTD dataset ISTD is proposed in (Wang, Li, and Yang 2018), which is the first public benchmark that could be used to train shadow detection and removal. SRD dataset SRD is proposed in (Qu et al. 2017), containing 2680 and 408 pairs of images for training and testing respectively. |
| Dataset Splits | No | This dataset has been divided into 1330 triplets for training and 540 triplets for testing. SRD does not provide masks, we employ the public SRD shadow masks from (Cun, Pun, and Shi 2020). |
| Hardware Specification | Yes | We implement our network in the Py Torch framework on the PC with a single NVIDIA Ge Force GTX 1080Ti GPU. |
| Software Dependencies | No | We implement our network in the Py Torch framework on the PC with a single NVIDIA Ge Force GTX 1080Ti GPU. In the training phase, we adopt the Adam optimizer (Kingma and Ba 2014) with a batch size of 2 and the patch size of 256 256. The initial learning rate is 5 10 5 and changes with Cosine Annealing scheme (Loshchilov and Hutter 2016). |
| Experiment Setup | Yes | In the training phase, we adopt the Adam optimizer (Kingma and Ba 2014) with a batch size of 2 and the patch size of 256 256. The initial learning rate is 5 10 5 and changes with Cosine Annealing scheme (Loshchilov and Hutter 2016). The CNNs parameters are randomly initialized and the model converges well after 150 epochs. For the hyperparameters, the weights (ηA, β, λ) in Equations (14) are initialized as 0.01 and these parameters can be automatically updated during the training phase in an end-to-end manner. The maximum iteration number K is empirically set to 4 as a trade-off between speed and accuracy. |