R2Net custom dataset

Overview This dataset contains a custom collection of images specifically designed and pre-processed to evaluate the performance of the Reversible Regression Network (R2-Net) in the context of block-based image compression. Context This dataset was created as part of the Bachelor's Thesis in Computer Engineering (TFG) titled "Desenvolupament en CUDA d'una tècnica de compressió basada en IA" at the Universitat Autònoma de Barcelona (UAB). The main goal of the project is to accelerate the R2-Net architecture overcoming the DPCM sequential bottleneck using parallelization strategies (Wavefront and Parallel) on CUDA GPUs. Data Sources & Licensing This dataset is a curated collection derived from public and open-source images to ensure reproducibility. The original sources include: Kodak Image Dataset: Widely used in the image compression literature (unrestricted use for research purposes). Wikimedia Commons: Images sourced under Creative Commons (CC0 or CC-BY) or Public Domain licenses. Note: This dataset is distributed under a Creative Commons Attribution (CC-BY 4.0) license for the compiled collection and preprocessing effort, respecting the original licenses of the source materials. Data Pre-processing To meet the strict mathematical requirements of the R2-Net algorithm, all images in this dataset have been subjected to the following preprocessing pipeline: Converted to grayscale (single color channel). Resized/Cropped to ensure dimensions are equal to the original R2Net article tests. Content The compressed archive contains 18 images in .raw format. Dataset Structure and Classes The following table summarizes the distribution and resolution of the images included in this collection, categorized by their classification criteria: Class Resolution Content Description Class A 2560 x 1600 Complex high-resolution city and crowd scenes. Class B 1920 x 1080 Detailed architectural and interior structures. Class C 832 x 480 Sports, animals, and natural landscapes. Class D 416 x 240 Close-up textures and small-scale patterns. Class E 1280 x 720 Portrait and park environments.