Intrinsic Single-Image HDR Reconstruction

TL;DR

Problem: Reconstructing HDR from a single LDR image is challenging due to saturation and nonlinear camera processing. Approach: A physically-mounded intrinsic HDR pipeline decomposes the image into shading $S$ and albedo $A$, learning separate HDR extensions in the shading domain via $D_H$ and in the albedo domain via $A_H$, with an additional refinement stage; shading uses the inverse domain $D_L = 1/(S_L+1)$ to handle long-tail illumination, and a soft-albedo guidance mask aids color recovery. Contributions: (i) a modular two-subtask HDR framework in the intrinsic domain, (ii) tailored loss terms on implied intrinsic components and an inverse-HDR representation, (iii) a dedicated refinement module, and (iv) strong quantitative and qualitative results across multiple in-the-wild datasets. Significance: enables high-fidelity recovery of bright-luminance details and accurate color from a single image, enhancing HDR display, editing, and downstream computational photography applications.

Abstract

The low dynamic range (LDR) of common cameras fails to capture the rich contrast in natural scenes, resulting in loss of color and details in saturated pixels. Reconstructing the high dynamic range (HDR) of luminance present in the scene from single LDR photographs is an important task with many applications in computational photography and realistic display of images. The HDR reconstruction task aims to infer the lost details using the context present in the scene, requiring neural networks to understand high-level geometric and illumination cues. This makes it challenging for data-driven algorithms to generate accurate and high-resolution results. In this work, we introduce a physically-inspired remodeling of the HDR reconstruction problem in the intrinsic domain. The intrinsic model allows us to train separate networks to extend the dynamic range in the shading domain and to recover lost color details in the albedo domain. We show that dividing the problem into two simpler sub-tasks improves performance in a wide variety of photographs.

Figures (8)

• Figure 1: In this work, we propose a physically-inspired HDR reconstruction pipeline by conducting reconstruction individually on intrinsic components. Our final reconstruction, as a result, is effective in recovering the high-luminance details and the lost colors.
• Figure 2: The in-camera pipeline modeling LDR image formation is shown at the top. Liu et al. liu2020single propose the inverse pipeline shown on the bottom, addressing each type of information loss individually in three steps. In this work, we use the pre-trained networks of Liu et al. liu2020single for the dequantization and linearization tasks (shown in pink) and focus on developing an intrinsic model for the dynamic range reconstruction problem highlighted in teal.
• Figure 3: We show the different effects of clipping on the albedo and the shading for a tonemapped HDR image and its LDR counterpart. When clipping the image, the LDR shading loses information compared to HDR. We scale the intensity of both shadings to show the effect. At the same time, the LDR albedo becomes saturated.
• Figure 4: Our HDR reconstruction pipeline starts with the intrinsic decomposition of the image using an off-the-shelf method careaga2023intrinsic. The LDR reconstruction is then done individually on shading and albedo layers, which are then combined in a final refinement stage. The input LDR image is provided to each network shown in blue.
• Figure 5: We show the difference between the shading representation and the L-channel from LUV. The shading correlates well with the geometry, while the L-channel includes texture information as shown in the green insets.