Mono4DGS-HDR: High Dynamic Range 4D Gaussian Splatting from Alternating-exposure Monocular Videos
Jinfeng Liu, Lingtong Kong, Mi Zhou, Jinwen Chen, Dan Xu
TL;DR
Mono4DGS-HDR is a novel system that reconstructs high dynamic range 4D scenes from unposed, alternating-exposure monocular videos, using a two-stage Gaussian Splatting approach to improve HDR rendering and consistency.
Contribution
The paper introduces the first method for HDR 4D scene reconstruction from monocular videos without camera poses, utilizing a two-stage Gaussian Splatting framework and a new HDR video benchmark.
Findings
Outperforms existing methods in rendering quality
Achieves faster reconstruction speeds
Provides robust HDR video reconstruction without camera poses
Abstract
We introduce Mono4DGS-HDR, the first system for reconstructing renderable 4D high dynamic range (HDR) scenes from unposed monocular low dynamic range (LDR) videos captured with alternating exposures. To tackle such a challenging problem, we present a unified framework with two-stage optimization approach based on Gaussian Splatting. The first stage learns a video HDR Gaussian representation in orthographic camera coordinate space, eliminating the need for camera poses and enabling robust initial HDR video reconstruction. The second stage transforms video Gaussians into world space and jointly refines the world Gaussians with camera poses. Furthermore, we propose a temporal luminance regularization strategy to enhance the temporal consistency of the HDR appearance. Since our task has not been studied before, we construct a new evaluation benchmark using publicly available datasets for…
Peer Reviews
Decision·ICLR 2026 Poster
1. The problem setting is well defined and motivated. 2. The paper is well written and clear. 3. The evaluations done are adequate, both quantitatively and qualitatively. 4. The paper comprehensively ablates all the design features showing the importance/visual effect of each modification.
1. How does the optmization/loss curves look like with that many losses? Would it be possible to show the curves? 2. Are all scenes at 24-30fps? Have the authors tried any more challenging settings like faster motion (ex - moving cars for autonomous driving applications?), non-lambertian surfaces etc? Those would be nice to haves but not necessary of course.
1. This paper presents the first system to address 4D HDR reconstruction from unposed, single-camera, alternating-exposure LDR videos. 2. The proposed two-stage optimization is effective. 3. HDR-4DGS effectively handles varying brightness across frames, which would break conventional photometric reprojection losses. 4. The paper constructs a new benchmark for HDR video reconstruction including real and synthetic scenes.
1. Although HDR reconstruction is the core contribution, the paper primarily evaluates PSNR/SSIM on tone-mapped images. No HDR-specific metrics (e.g., PQ-PSNR, HDR-VDP). 2. While quantitative results are extensive, the paper provides limited qualitative discussion on typical failure cases. 3. The approach has not been evaluated on low-light, reflective, or transparent surfaces, which may limit applicability in certain real-world conditions.
1. Explicitly parameterizing the motion of Gaussians not only improves rendering quality but also maintains a relatively fast rendering speed. 2. The invariance of 2D Gaussian covariance serves as a simple yet effective tool introduced by the authors, which is validated through ablation studies in the paper. 3. The entire paper is clear and easy to understand.
1. The division between dynamic and static regions relies on epipolar error maps, so the final results are heavily influenced by them. 2. The selection of dynamic Gaussians depends on threshold settings, which reduces the generalizability of the pipeline, as determining appropriate thresholds for each scene is not straightforward.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsImage Enhancement Techniques · Advanced Vision and Imaging · Generative Adversarial Networks and Image Synthesis