Casual3DHDR: High Dynamic Range 3D Gaussian Splatting from Casually Captured Videos

ACM Multimedia 2025

Shucheng Gong1,2*    Lingzhe Zhao1*    Wenpu Li1*    Hong Xie2†    Yin Zhang1,4    Shiyu Zhao1    Peidong Liu1†

*equal contribution    † denotes corresponding author.

1Westlake University    2Wuhan University    3Zhejiang University   

arXiv Code

Given a casually captured video with auto exposure, camera motion blur, and significant exposure time changes, we train 3DGS to reconstruct a sharp HDR scene.

After reconstructing the 3D HDR scene, we can render sharp LDR videos (for standard monitor display) with any given exposure time and camera trajectory.

Keywords

High dynamic range, Motion blur, 3D Gaussian Splatting, Novel view synthesis, Casual video, Exposure time estimation, Camera response function, Pose optimization, Camera trajectory, B-Spline curve


Teaser

overview

a) Our method can reconstruct 3D HDR scenes from videos casually captured with auto-exposure enabled.

b) Our approach achieves superior rendering quality compared to methods like Gaussian-W and HDR-Plenoxels.

c) After 3D HDR reconstruction, we can not only synthesize novel view, but also perform various downstream tasks, such as 1) HDR exposure editing, 2) Image deblurring.


Abstract

Photo-realistic novel view synthesis from multi-view images, such as neural radiance field (NeRF) and 3D Gaussian Splatting (3DGS), has gained significant attention for its superior performance.
However, most existing methods rely on low dynamic range (LDR) images, limiting their ability to capture detailed scenes in high-contrast environments. While some prior works address high dynamic range (HDR) scene reconstruction, they typically require multi-view sharp images with varying exposure times captured at fixed camera positions, which is time-consuming and impractical.
To make data acquisition more flexible, we propose Casual3DHDR, a robust one-stage method that reconstructs 3D HDR scenes from casually-captured auto-exposure (AE) videos, even under severe motion blur and unknown, varying exposure times.
Casual3DHDR integrates a continuous-time camera trajectory into a unified physical imaging model, jointly optimizing exposure times, camera trajectory, and the camera response function (CRF).
Extensive experiments on synthetic and real-world datasets demonstrate that Casual3DHDR outperforms existing methods in robustness and rendering quality.


Pipeline

overview

Given a casually captured video with auto exposure, camera motion blur, and significant exposure time changes, we train 3DGS to reconstruct an HDR scene.

We design a unified model based on the physical image formation process, integrating camera motion blur and exposure-induced brightness variations.

This allows for the joint estimation of camera motion, exposure time, and camera response curve while reconstructing the HDR scene.

After training, our method can sharpen the train images and render HDR and LDR images from specified poses.


Estimated Exposure Times

exposure time estimation
exposure time estimation 2

The figures show the comparison between the jointly-optimized exposure times and the ground truth exposure times for each training image in scenes Toufu-vicon (left) and Girls-vicon (right). The results are scaled uniformly, and it can be observed that the estimated exposure times closely follow the trend of the ground truth exposure times.


Acknowledgements

We specially thank Xiang Liu for his kind and valuable suggestions during the writing of this paper.


BibTeX

@inproceedings{gong2025casual3dhdr,
              title={{Casual3DHDR: High Dynamic Range 3D Gaussian Splatting from Casually Captured Videos}},
              author={Gong, Shucheng and Zhao, Lingzhe and Li, Wenpu and Xie, Hong and Zhang, Yin and Zhao, Shiyu and Liu, Peidong},
              booktitle={In Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25)},
              year={2025},
            }