ICAA17K

Dataset Overview

Dataset Name

• ICAA17K

Dataset Description

• ICAA17K is designed for image color aesthetic assessment (ICAA) tasks and is currently the largest and most densely annotated ICAA dataset, encompassing diverse color types and acquisition devices.

Dataset Features

• To remedy the lack of color annotations in existing IAA datasets, ICAA17K provides detailed color labels, avoiding bias toward single colors (e.g., black‑white).
• The dataset includes richer color types and combinations, reducing over‑concentration on any single hue.

Dataset Download

• Available via Google Drive or Baidu Drive.

Models and Methods

Model Name

• Delegate Transformer

Model Description

• The Delegate Transformer learns to segment the color space through specialized deformable attention rather than static pixel values, thereby capturing spatial color information.
• The model assigns different attention weights based on color importance, enhancing fine‑grained color perception.

Model Weights

• Currently, due to project constraints, model weights are not publicly released, but training code is available for users to train independently.

Benchmarking

Benchmark Description

• Based on the ICAA17K dataset, a large benchmark comprising 15 methods for image color aesthetic evaluation has been released, representing the most comprehensive ICAA benchmark to date.

Benchmark Datasets

• Evaluations are conducted on both the SPAQ and ICAA17K datasets.

Environment and Execution

Environment Requirements

• Install required packages such as pandas, nni, requests, torchvision, numpy, scipy, tqdm, torch, scikit_learn, tensorboardX, etc.

Execution Guide

• Prior to training or testing, load pretrained weights from the provided link or train them yourself.
• Use the nni tool for training and testing, or modify the code to run without nni as needed.