Abstract

Photometric invariance is vital in many computer vision tasks. Achieving robustness to photometric variations in imaging conditions requires the collection of accurate and sufficient ground-truth data for training. However, this task proves to be challenging, and as a result the reliance on synthetic data compromises the capacity of the model to adapt and generalize well to real-world situations. To address this issue, we move beyond purely data-driven paradigms by introducing physics priors as photometric invariants applicable across diverse models, datasets, and vision tasks. Specifically, we extend beyond task-specific physics priors and present a novel framework that systematically analyzes their effectiveness, thereby enabling a rigorous assessment of photometric-invariant physics priors across various vision tasks, such as semantic segmentation, intrinsic image decomposition (IID), color constancy, and image classification. We propose a physics-based self-supervised learning framework that extracts photometric-invariant features from unlabeled real-world images. Our approach integrates multiple physics priors and uncertainty modeling into a U-Net architecture, enabling robust, model-agnostic, and task-agnostic feature representations while facilitating efficient transfer learning under limited data regimes. Furthermore, we introduce IDTransformer, a transformer-based model incorporating photometric-invariant attention. In contrast to prior methods relying on hand-crafted priors or purely data-driven learning, IDTransformer captures reflectance transitions and clusters similar reflectance regions independently of spatial arrangement. By leveraging illumination- and geometry-invariant attention for reflectance mapping and geometry-variant attention for shading estimation, it achieves competitive performance with minimal training data.