SiamTrans: Zero-Shot Multi-Frame Image Restoration with Pre-trained Siamese Transformers
Authors: Lin Liu, Shanxin Yuan, Jianzhuang Liu, Xin Guo, Youliang Yan, Qi Tian1747-1755
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we show ablation study and comparison with state-of-the-art methods. Our algorithm is implemented on a NVIDIA Tesla V100 GPU in Py Torch. |
| Researcher Affiliation | Collaboration | Lin Liu1, Shanxin Yuan2*, Jianzhuang Liu2, Xin Guo1, Youliang Yan2, Qi Tian3 1EEIS Department, University of Science and Technology of China 2Huawei Noah s Ark Lab 3Huawei Cloud BU {ll0825,willing}mail.ustc.edu.cn {shanxin.yuan, liu.jianzhuang, yanyouliang, tian.qi1}@huawei.com |
| Pseudocode | No | The paper describes the method and uses mathematical formulations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing the source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | The pre-training task is denoising with the Place365 dataset (Zhou et al. 2017). ... Since there is no existing short-sequence deraining dataset, we build our multi-frame deraining test set through extracting adjacent frames from the NTURain dataset (Chen et al. 2018)... The training set for the compared supervised methods is Rain100L (Yang et al. 2017)... |
| Dataset Splits | No | The paper describes training and testing sets for various datasets, often for 'compared supervised methods'. However, it does not explicitly specify a validation dataset split or percentages used for its own model's development or tuning in a traditional sense. |
| Hardware Specification | Yes | Our algorithm is implemented on a NVIDIA Tesla V100 GPU in Py Torch. |
| Software Dependencies | No | The paper states: 'Our algorithm is implemented on a NVIDIA Tesla V100 GPU in Py Torch.' While PyTorch is mentioned, no specific version number is provided for it or any other software dependencies. |
| Experiment Setup | Yes | In both the pre-training and zero-shot restoration, the batch size is set to 1 and the initial learning rate is 1 10 5. The algorithm runs for 20 epochs and 20 iterations for the pre-training and the hard patch refinement, respectively. For zero-shot restoration, it takes 200, 500, and 1000 iterations for demoir eing, desnowing, and deraining, respectively. The λ and α in Eqn. 8 and Eqn. 9/Eqn. 10 are empirically set to 5 and 0.9 respectively. |