IllumiCAM: Grad-CAM Multi-Illuminant Detection & Selective White Balance

Overview

IllumiCAM implements multiple CNN architectures for illuminant estimation, classifying scene illuminants into five categories (Very Warm, Warm, Neutral, Cool, Very Cool) and using Class Activation Maps to visualize spatial attention. The key idea is to turn those CAMs into a spatially-aware white balance correction — using CAM-guided masks to apply selective correction across multi-illuminant scenes — evaluated against the LSMI (Localized Spatially Mixed Illuminant) test set.

Models & methods

• IlluminantCNN — a custom 5-layer CNN with BatchNorm and global max pooling; best test accuracy 84.49%.
• Confidence-Weighted CNN — FC4-inspired, with learned spatial confidence weighting instead of fixed pooling.
• ColorConstancyCNN — an AlexNet-based architecture from the color constancy literature.
• IllumiCam3 — global average pooling for cleaner CAM interpretability.
• CAM methods compared: Grad-CAM, Grad-CAM++, and Score-CAM, evaluated against ground-truth masks via IoU, DICE, and MAE.