Invited Talks

High Dynamic Range (HDR) images can capture and display much richer appearance information than Low Dynamic Range (LDR) images, playing a vital role in image representation and visualization. In this talk, we discuss the neural reconstruction of HDR content from single and multiple LDR images, as well as specific aspects of high-fidelity reproduction of HDR content on LDR displays.

Initially, we introduce a generative model for HDR images, trained exclusively on diverse LDR image datasets in a fully unsupervised fashion. Leveraging exposure variations within these datasets, which implicitly contain information about the underlying distribution of HDR images, we propose the GlowGAN approach. This method enables the generation of extensive HDR image datasets and the reconstruction of HDR content from individual LDR images, aligning with the unsupervised inverse tone mapping (ITM) task, which generates believable details for overexposed areas. Secondly, we propose a novel approach that combines implicit neural representations with an expressive camera model to reconstruct an all-in-focus HDR image from an image stack. By skillfully mixing components of the stack across different focal distances, apertures, and exposure times, we achieve better results than those obtained with commonly used homogeneous focal or exposure stacks, which are more redundant for neural reconstruction. Finally, we focus on faithfully reproducing material gloss–a challenging task given the limited dynamic range, peak luminance, and 3D capabilities of display devices. Through psychophysical experiments on a wide range of 3D printed samples and their corresponding HDR photographs, we quantify changes in gloss appearance due to display limitations. These measurements inform adjustments to material reflectance parameters within a rendering system, aiming to improve the fidelity of gloss representation between real objects and their visualization on display devices.