Stochastic Window Transformer for Image Restoration

Authors: Jie Xiao, Xueyang Fu, Feng Wu, Zheng-Jun Zha

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments validate the stochastic window strategy consistently improves performance on various image restoration tasks (deraining, denoising and deblurring) by significant margins.
Researcher Affiliation Academia Jie Xiao, Xueyang Fu , Feng Wu, Zheng-Jun Zha University of Science and Technology of China, Hefei, China ustchbxj@mail.ustc.edu.cn, {xyfu,fengwu,zhazj}@ustc.edu.cn
Pseudocode No The paper describes its methods using mathematical equations and textual explanations (e.g., Equations (1), (2), (4), (5)) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The code is available at https://github.com/jiexiaou/Stoformer.
Open Datasets Yes All these methods are evaluated on SPA-Data [51]... We conduct denoising experiments on the additive white Gaussian Noise benchmark datasets, which include Set12 [66], BSD68 [37], Urban100 [19], Kodak24 [12] and Mc Master [71]... We also perform deblurring experiments on the benchmark datasets (Go Pro [39] and HIDE [45]).
Dataset Splits No The training samples are augmented by the horizontal flipping and rotation of 90 , 180 , or 270 . Please refer to the supplemental material for task-specific settings. The paper states 'Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See supplemental material.' The main text does not explicitly detail the train/validation/test splits, only referencing the supplemental material.
Hardware Specification No The paper states 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See supplemental material.' indicating that specific hardware details are not provided in the main text.
Software Dependencies No The paper mentions using 'Charbonnier loss' and 'Adam optimizer' but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes Loss Function. The loss function adopted for training is the Charbonnier loss [1], whose mathematical expression is: L(I , I) = p ||I I||2 + ϵ2, (6) where I and I are the restored and ground-truth image respectively. The constant ϵ is empirically set to 10 3. Training Detail. Stoformer employs a four-level encoder-decoder structure. The numbers of Sto Block are {1, 2, 8, 8} for level-1 to level-4 of Encoder and the blocks for Decoder are mirrored. The number of channel is set to 32 and the window size is 8 8. We train the network with Adam optimizer (β1 = 0.9, β2 = 0.999) with the initial learning rate 3 10 4 gradually reduced to 1 10 6 with the cosine annealing. The training samples are augmented by the horizontal flipping and rotation of 90 , 180 , or 270 . Please refer to the supplemental material for task-specific settings.