Characterising and dissecting human perception of scene complexity

Area of Science:

• Cognitive Psychology
• Computer Vision
• Neuroscience

Background:

• Human ability to assess visual complexity is effortless but poorly understood.
• Natural scenes present a constant challenge for complexity perception research.
• Existing models often overlook the interplay between low-level features and semantics.

Purpose of the Study:

• To investigate the factors influencing human perception of scene complexity.
• To develop and validate computational models for predicting perceived complexity.
• To propose a novel framework explaining dual information processing in visual complexity.

Main Methods:

• Creation of two novel datasets (VISC-C and VISC-CI) with human complexity annotations.
• Analysis of perceptual features (clutter, symmetry, entropy, openness) using hierarchical regression.
• Evaluation of neural network models and validation against a large complexity dataset.
• Dissection of best-performing neural networks to understand feature extraction.

Main Results:

• Both global image properties and semantic features significantly contribute to complexity perception.
• Combining perceptual features with semantic network output improved variance explanation.
• Neural networks learn to extract both low-level and high-level scene details for complexity prediction.

Conclusions:

• Human complexity perception relies on a "dual information" framework, integrating low-level and high-level processing.
• Novel datasets and analyses provide a foundation for future research in visual complexity.
• Computational models can effectively predict human complexity perception by incorporating semantic understanding.