Transformer-Based Selective Super-resolution for Efficient Image Refinement

Authors: Tianyi Zhang, Kishore Kasichainula, Yaoxin Zhuo, Baoxin Li, Jae-Sun Seo, Yu Cao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three datasets demonstrate the efficiency and robust performance of our approach for super-resolution.
Researcher Affiliation Academia 1University of Minnesota 2Arizona State University 3Cornell Tech
Pseudocode Yes We outline the precise algorithm in Algorithm 1.
Open Source Code No The paper does not provide concrete access to source code (specific repository link, explicit code release statement, or code in supplementary materials) for the methodology described.
Open Datasets Yes We evaluate the effectiveness of our SSR model on three datasets, BDD100K (Yu et al. 2020), COCO 2017 (Lin et al. 2014) and MSRA10K (Cheng et al. 2015).
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) needed to reproduce the data partitioning for all datasets. While COCO2017 mentions training and validation images, it does not state how the authors applied this or similar splits across all datasets used in their experiments.
Hardware Specification Yes All experiments are conducted for 50 epochs on two Linux servers, each equipped with two NVIDIA A6000 GPUs.
Software Dependencies No The paper mentions 'YOLOv8' but does not provide a specific version number, nor does it list other software dependencies with version numbers.
Experiment Setup Yes The embedding dimension of the Tile Selection (TS) module is 96 while it s 180 for Tile Refinement (TR) module. We set the learning rates to 0.00001. Each TL utilizes a depth of 2, a window size of 7, and 3 attention heads. We employ a patch size of 2 for embedding, which corresponds to tile sizes of 16 16, 32 32, and 64 64, yielding tile labels of 4 4, 8 8, and 16 16 respectively. The weight parameter α for the loss function is set to 1. The number of RTB is 6. All experiments are conducted for 50 epochs