SeHDR: Single-Exposure HDR Novel View Synthesis Via 3D Gaussian Bracketing

NVS methods, our approach is applicable to any stan- 104 dard multi-view datasets seamlessly, eliminating the need 105 for specialized multi-exposure capture setups and thereby 106 broadening practical applicability. 107 • We introduce exposure-bracketed 3D Gaussians, ex- 108 tending exposure-bracketing principles to 3D space to 109 synthesize multi-exposure 3D Gaussians for fusion into 110 HDR 3D Gaussians. To enable this fusion, we further 111 design a Neural Exposure Fusion (NeEF) method, which 112 operates within the SH space, to effectively reconstruct 113 HDR radiance fields.

114 • Extensive comparative experiments against state-of-the-115 art HDR-NVS methods and baselines demonstrate that 116 our approach achieves superior performances in dynamic 117 range expansion and photorealism.

118 2. Related Works 119 HDR-NVS. Recent advances have focused on directly re-120 constructing HDR scenes from multi-view LDR input im-121 ages. Huang et al. [9] propose the first HDR-NVS method, 122 HDR-NeRF, that integrates a simplified physical imaging 123 process into the NeRF [18] framework. Unlike standard 124 NeRF that models density and LDR color along a ray, 125 HDR-NeRF encodes density and unbounded scene radiance 126 (ranging from 0 to +→) along each ray, employing a learn-127 able tone-mapping MLP to project radiance back into LDR 128 cies. To address these limitations, we propose SEHDR, a 137 method for reconstructing HDR scenes directly from single-138 exposure LDR images, which is able to take any standard 139 multi-view dataset as inputs.

140 HDR Image Reconstruction. Conventional methods for 141 reconstructing high dynamic range (HDR) images typically 142 rely on two approaches: estimating an intermediate camera 143 response function (CRF) from multiple exposure low dy-144 namic range (LDR) images [2, 6], or directly fusing brack-145 eted LDR images with identical camera poses but varying 146 exposures into HDR representations using intensity-prior-147 guided weight maps [17]. These methods inherently require 148 multiple exposure inputs to generate HDR outputs. Recent 149 advances in deep learning have enabled subsequent stud-150 ies to explore single-exposure LDR-to-HDR reconstruction 151 through convolutional neural networks (CNNs) [4, 5, 15], 152 generative adversarial networks (GANs) [13], and deep 153 residual networks (ResNets) [3]. 154 Integrating pre-trained single-image HDR reconstruc-155 tion techniques with Gaussian splatting may be a plausible 156 solution for the HDR-NVS problem. However, in this work, 157 we show that such strategies often prove ineffective due 158 to multi-view inconsistencies introduced by single-image 159 methods, resulting in obvious floater and blurry artifacts in 160 the rendered novel view images (cf ., Fig. 1 (d)). 161 Novel View Synthesis (NVS). Novel View Synthesis 162 (NVS) aims to generate photorealistic images of a scene 163 Figure 1. Given a set of multi-view images with single exposure (a), our method renders a photorealistic high quality novel view image with accurate color and fine details (b), while existing multi-exposure based HDRGS [2] method fails to reconstruct HDR colors from a single-exposure (c). Combining pre-trained HDR image reconstruction methods with 3DGS tends to fail to render accurate details and produce blurs and floaters, e.g., on the stonewall region (d). Note that both SEHDR and HDRGS are trained without HDR ground truth. We use PhotoMatix Pro [27] with the same settings to tone-map HDR results for visualization.