Dataset for DCCN

Dataset Description This archive contains training materials and a minimal dataset for image aesthetic quality assessment based on two widely-used datasets: AVA and CUHK-PQ. 1. AVA Dataset (Aesthetic Visual Analysis) The AVA dataset consists of approximately 250,000 images, each annotated with a distribution of aesthetic scores (from 1 to 10). In our experiments, we used the complete AVA dataset for training to achieve optimal performance. However, due to the large file size (~30GB), it is not included in this archive. Instead, we provide a minimal version of the dataset with around 30,000 selected images and a merged CSV file, which can be used for lightweight training and testing. Please note that models trained on this smaller subset may perform worse than those trained on the full dataset. You can download the full AVA dataset using this tool:https://github.com/imfing/ava_downloader 2. CUHK-PQ Dataset The CUHK-PQ dataset is organized into two folders: high/ – High aesthetic quality images low/ – Low aesthetic quality images The folder structure itself serves as ground truth labels, making the dataset straightforward to use in binary classification tasks. In this project, the CUHK-PQ dataset was used to train and evaluate a binary aesthetic classifier. 3. Code The code package contains implementations for training and evaluating models on both the AVA and CUHK-PQ datasets.This code enables reproduction of the experimental results presented in this study. Archive Contents minimal_dataset_for_ava.csv – Combined label file for AVA subset (with train/val/test split) AVA/ – Folder with selected AVA training images README.md – Dataset usage instructions A sample image showing the training result visualization result_ava.png – Sample visualization of model performance on the AVA dataset result_pq.png – Sample visualization of model performance on the CUHK-PQ dataset Source code for training and evaluating models on AVA and CUHK-PQ datasets