IC9600: A Benchmark Dataset for Automatic Image Complexity Assessment.

Area of Science:

• Computer Vision
• Human Perception
• Machine Learning

Background:

• Image complexity (IC) is crucial for visual understanding but challenging to evaluate due to subjectivity and semantic dependency.
• Existing methods for IC assessment are limited, hindering research in this area.
• The diversity of real-world images further complicates objective IC evaluation.

Purpose of the Study:

• To create the first large-scale dataset for image complexity (IC) assessment.
• To develop a weakly supervised model for predicting IC scores and complexity density maps.
• To demonstrate the utility of IC in enhancing various computer vision tasks.

Main Methods:

• Construction of a novel dataset comprising 9,600 diverse images (abstract, paintings, real-world scenes).
• Annotation of images by 17 human contributors to capture subjective complexity.
• Development of a weakly supervised base model for IC score prediction and complexity density mapping.

Main Results:

• The developed IC dataset is the largest to date, featuring high-quality annotations.
• The base model effectively predicts IC scores, achieving a high correlation (Pearson coefficient: 0.949) with human perception.
• Empirical validation shows that IC information boosts the performance of multiple computer vision tasks.

Conclusions:

• The IC9600 dataset provides a valuable resource for advancing research in image complexity assessment.
• Weakly supervised learning is effective for predicting image complexity and density maps.
• Incorporating image complexity offers significant benefits for a broad spectrum of computer vision applications.