Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
CascadedGaze: Efficiency in Global Context Extraction for Image Restoration
Authors: Amirhosein Ghasemabadi, Muhammad Kamran Janjua, Mohammad Salameh, CHUNHUA ZHOU, Fengyu Sun, Di Niu
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results show that our computationally efficient approach performs competitively to a range of state-of-the-art methods on synthetic image denoising and single image deblurring tasks, and pushes the performance boundary further on the real image denoising task. |
| Researcher Affiliation | Collaboration | Amirhosein Ghasemabadi1,2, Muhammad Kamran Janjua 2, Mohammad Salameh 2, Chunhua Zhou3, Fengyu Sun3, Di Niu1 1Dept. ECE, University of Alberta, Canada, 2Huawei Technologies, Canada, 3Huawei Kirin Solution, China |
| Pseudocode | No | The paper describes its methodology in Section 3 and provides an architecture diagram in Figure 2, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We release the code at the following link: https://github.com/Ascend-Research/Cascaded Gaze. |
| Open Datasets | Yes | For image denoising, we train our models on both synthetic benchmark datasets (Gaussian image denoising) and the real-world noise dataset (real image denoising). The Smartphone Image Denoising Dataset (SIDD) (Abdelhamed et al., 2018) is a real-world noise dataset... synthetic benchmark datasets are generated with additive white Gaussian noise on BSD68 (Martin et al., 2001), Urban100 (Huang et al., 2015), Kodak24 (Franzen, 1999) and Mc Master (Zhang et al., 2011). For image motion deblurring, we employ the Go Pro dataset (Nah et al., 2017) as the training data. |
| Dataset Splits | Yes | During training for real denoising and motion deblurring experiments, we set the image patch size to 256 256. For Gaussian denoising, we follow (Zamir et al., 2022) s progressive training configuration and start with the patch size of 160 and increase it to 192, 256, 320, and 384 during training. |
| Hardware Specification | Yes | All of our models are implemented in the Py Torch library, trained on 8 NVIDIA Tesla v100 PCIe 32 GB GPUs. For inference, we utilize a single GPU. |
| Software Dependencies | No | The paper states, "All of our models are implemented in the Py Torch library," but it does not provide a specific version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We train the models in all tasks for 400K iterations, with Adam W as the optimizer (β1 = 0.9, β2 = 0.9), and minimize the negative PSNR loss function (i.e., maximize the PSNR). We use a cosine annealing scheduler that starts with the learning rate of 1e 3 and decays to 1e 7 throughout learning. All of our models are implemented in the Py Torch library, trained on 8 NVIDIA Tesla v100 PCIe 32 GB GPUs. For inference, we utilize a single GPU. During training for real denoising and motion deblurring experiments, we set the image patch size to 256 256. For Gaussian denoising, we follow (Zamir et al., 2022) s progressive training configuration and start with the patch size of 160 and increase it to 192, 256, 320, and 384 during training. |