Bridging the Domain Gap: A Physics-Aware Synthetic Data Pipeline for Multi-Spectral Facial Analysis

Current computer vision models struggle with domain generalization, particularly in non-visible spectra (e.g., thermal imaging) where labeled data is scarce. This paper proposes a unified framework that leverages intrinsic image decomposition (separating albedo from illumination) and physically based rendering to generate large-scale, photorealistic synthetic datasets. By simulating accurate optical lens characteristics and thermal heat signatures on 3D virtual humans, we demonstrate that synthetic data can successfully train state-of-the-art CNNs for tasks like gender classification in the thermal domain, achieving performance parity with real-world data while preserving privacy and eliminating dataset bias. Our contributions include a multi-modal synthesis engine, an SVR-optimized optical augmentation module, and an inverse-to-forward loop for capturing realistic skin textures. Extensive mathematical formalization and experimental validation confirm the efficacy of our approach.