See More Details: Efficient Image Super-Resolution by Experts Mining

Authors: Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, Radu Timofte

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4. Experiments Datasets and Evaluation. Following the SR literature (Liang et al., 2021; Chen et al., 2023), we utilize DIV2K (Agustsson & Timofte, 2017) and Flickr2K (Lim et al., 2017) datasets for training. We produce LR images using bicubic downscaling of HR images. When testing our method, we assess its performance on canonical benchmark datasets for SR Set5 (Bevilacqua et al., 2012), Set14 (Zeyde et al., 2010), BSD100 (Martin et al., 2001), Urban100 (Huang et al., 2015) and Manga109 (Matsui et al., 2017). We calculate PSNR and SSIM results on the Ychannel from the YCb Cr color space.
Researcher Affiliation Academia 1Computer Vision Lab, CAIDAS & IFI, University of W urzburg, Germany 2AI Institute, Shanghai Jiao Tong University, China 3Computer Vision Lab, ETH Zurich, Switzerland.
Pseudocode Yes Algorithm 1 Mixture of Low-Rank Experise
Open Source Code Yes The source codes will be publicly made available at https://github. com/eduardzamfir/seemoredetails
Open Datasets Yes Following the SR literature (Liang et al., 2021; Chen et al., 2023), we utilize DIV2K (Agustsson & Timofte, 2017) and Flickr2K (Lim et al., 2017) datasets for training.
Dataset Splits No The paper does not explicitly state a training/validation/test split. It mentions training on DIV2K and Flickr2K and testing on Set5, Set14, BSD100, Urban100, and Manga109, which are standard test benchmarks, but no specific validation split from the training data is provided.
Hardware Specification Yes All experiments are conducted with the Py Torch framework on NVIDIA RTX 4090 GPUs.
Software Dependencies No The paper mentions 'PyTorch' and 'Adam optimizer' but does not specify their version numbers. Appendix A mentions 'Py Torch-based Basic SR' framework and 'fvcore Python package' but without version numbers.
Experiment Setup Yes Similar to (Sun et al., 2022; 2023), we minimize the L1-Norm between SR output and HR ground truth in the pixel and frequency domain using Adam (Kingma & Ba, 2017) optimizer for 500K iterations with a batch size of 32 and initial learning rate of 1 10 3 halving it at following milestones: [250K,400K,450K,475K]. The feature dimension and channel expansion factor in Gated FFN are set to 36 and 2, respectively. For all Mo RE sub-modules, we select an exponential growth of the channel dimensionality and choose in total of 3 experts. The kernel size in SEE is set to 11 11.