Color Constancy by Reweighting Image Feature Maps.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Computational color constancy aims to correct color casts caused by illumination.
• Existing deep learning models offer high accuracy but lack interpretability and can be computationally expensive.
• Interpretable models provide insights but may not match the performance of deep learning approaches.

Purpose of the Study:

• To propose a novel illuminant color estimation framework for computational color constancy.
• To integrate the representational power of deep learning with the interpretability of assumption-based models.
• To develop a computationally efficient model with high accuracy and provide uncertainty estimates.

Main Methods:

• A novel framework for illuminant color estimation is proposed.
• A feature map reweight unit (ReWU) is designed as a key building block.
• A confidence estimation branch is incorporated for simultaneous point and uncertainty estimation.

Main Results:

• The proposed framework achieves comparative accuracy on benchmark datasets against state-of-the-art deep learning models.
• The model demonstrates a more compact size and lower computational cost compared to existing methods.
• The confidence estimation branch provides useful uncertainty estimates for aggregation and multiple illumination estimation.

Conclusions:

• The novel framework effectively balances accuracy, interpretability, and computational efficiency in illuminant color estimation.
• The ReWU and confidence estimation branch contribute to improved performance and provide valuable uncertainty information.
• The availability of source code and datasets facilitates further research and application in computational color constancy.