HDR-GS: Efficient High Dynamic Range Novel View Synthesis at 1000x Speed via Gaussian Splatting

High dynamic range (HDR) novel view synthesis (NVS) aims to create photorealistic images from novel viewpoints using HDR imaging techniques. The rendered HDR images capture a wider range of brightness levels containing more details of the scene than normal low dynamic range (LDR) images. Existing HDR NVS methods are mainly based on NeRF. They suffer from long training time and slow inference speed. In this paper, we propose a new framework, High Dynamic Range Gaussian Splatting (HDR-GS), which can efficiently render novel HDR views and reconstruct LDR images with a user input exposure time. Specifically, we design a Dual Dynamic Range (DDR) Gaussian point cloud model that uses spherical harmonics to fit HDR color and employs an MLP-based tone-mapper to render LDR color. The HDR and LDR colors are then fed into two Parallel Differentiable Rasterization (PDR) processes to reconstruct HDR and LDR views. To establish the data foundation for the research of 3D Gaussian splatting-based methods in HDR NVS, we recalibrate the camera parameters and compute the initial positions for Gaussian point clouds. Experiments demonstrate that our HDR-GS surpasses the state-of-the-art NeRF-based method by 3.84 and 1.91 dB on LDR and HDR NVS while enjoying 1000x inference speed and only requiring 6.3% training time. Code and recalibrated data will be publicly available at https://github.com/caiyuanhao1998/HDR-GS . A brief video introduction of our work is available at https://youtu.be/wtU7Kcwe7ck

Paper

Similar papers

Peer review

Reviewer VwS86/10 · confidence 5/52024-07-06

Summary

The manuscript presents a novel approach to High Dynamic Range (HDR) novel view synthesis (NVS) by proposing a new framework called High Dynamic Range Gaussian Splatting (HDR-GS). The proposed HDR-GS framework addresses the limitations of existing HDR NVS methods, which are primarily based on Neural Radiance Fields (NeRF) and suffer from long training times and slow inference speeds.

Strengths

Comprehensive experiments demonstrate that HDR-GS outperforms state-of-the-art NeRF-based methods while achieving 1000× faster inference speed and requiring only 6.3% of the training time. The methodology is straightforward and well-organized, and the experimental results are impressive. The paper is logically structured and easy to follow.

Weaknesses

While the methodology is straightforward and the experimental results are strong, the authors could further highlight their innovations and provide detailed explanations of their model design and theoretical underpinnings. This would enhance the manuscript's persuasiveness.

Questions

1) The authors' method is simple and effective, and the significant speed improvement due to 3D Gaussian Splatting (3D-GS) is not surprising. However, it would be beneficial if the authors could provide an analysis of what specifically leads to the substantial performance difference between HDR-GS and HDR-NeRF. 2) The authors chose to model HDR color first using the initialized DDR Gaussian point cloud rather than LDR color. It would be helpful if the authors could elaborate on the rationale behind this choice and how it positively impacts the generation of HDR and LDR results.

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Section 3.2 could be omitted or condensed. The mathematical explanation of 3D-GS rendering and the snowballing process does not seem to be a primary innovation of this paper. Additionally, the emphasis on parallel HDR and LDR rendering and rasterization is somewhat unclear. The authors should clarify why parallel rendering is necessary for their approach, as it seems that similar results could be achieved without parallel rendering. Specifically, the necessity of parallel rendering for the subsequent loss calculation involving 𝐼𝑙 and 𝐼ℎ should be justified.

Reviewer tLXM7/10 · confidence 5/52024-07-06

Summary

This paper proposes a 3D Gaussian Splatting-based method, HDR-GS, for the high dynamic range novel view synthesis. To efficiently perform this task, a new Dual Dynamic Range Gaussian point cloud model is presented (in Section 3.1). This point cloud model has more attributes including the HDR color, exposure time, and a global-shared neural network functioning as a tone-mapper. Then in Section 3.2, two Parallel Differentiable Rasterization processes are designed to render the low and high dynamic range colors. Besides, in Section 3.3, the authors recalibrate the camera parameters for the real and synthetic multi-view HDR datasets to make the scene-level data suitable for 3D Gaussian Splatting-based algorithms.

Strengths

+ It is a good attempt to design the first 3DGS-based framework for the task of high dynamic range novel view synthesis. The core idea of assigning more attributes to the Gaussian point cloud model and rendering the high dynamic range and low dynamic range views in the parallel rasterization is novel and cool, which makes the proposed 3D Gaussian model multifunctional and have great practical values in photography, film making, etc. + The performance is superior. The running speed of the proposed HDR-GS is more than a thousand time that of the state-of-the-art NeRF-based method, HDR-NeRF. As compared in Table 1 of the main paper, previous NeRF-based methods suffer from the slow inference speed (< 0.2 fps). The proposed HDR-GS can not only infer at a much faster speed of 126 fps (>> 30 fps) but also surpass the SOTA method by large margins. These advantages enable the HDR-GS to capture and measure dynamic scenes, e.g., the camera on the robot, in real time. + The writing style is clear and easy to follow. I notice that the authors did not follow the mainstream introduction of the 3D Gaussian Splatting part, which I think is very confusing. Instead, the authors adopted the strategy of summarizing first and then dividing. They first introduced what attributes the Gaussian point cloud model contains, and then gradually introduced its working pipeline. This introduction order can give the readers an overall knowledge of what the Gaussian model is (its attributes) and thus helps the readers better understand the method. + The data re-calibration is a bonus. In my opinion, the biggest obstacle to researching the topic of high dynamic range novel view synthesis is the data issue. Although the multi-view synthetic and real datasets are collected, the normalized device coordinates and lack of SfM points for initialization cannot make the 3DGS work because of the severe blur and overfitting problems, especially for the unbounded scene-level reconstruction. How to re-calibrate the data is critical and executing the SfM algorithm is time-consuming.

Weaknesses

There are some minor issues: - In Line 124 – 131, the authors analyzed the advantages of using log tone-mapper than linear one from the point of view of training stability. Yet, my understanding is that the option of taking the logarithm can shorten the gaps between the training data samples, which makes the originally discrete data samples become more continuous. The processed data samples are easier to fit by neural networks. The exposure time in table 3d is an example. So, adding this analysis can help better explain the motivation of taking the logarithm instead of directly using the linear form. - In Eq (14), the authors just used part of images of a scene to re-calibrate the camera poses without explanation. My concern is why not use all of the images instead of just using the views under the same exposure time? Did you try that? It is interesting to know and analyze this result. - It would be better to add some legend or annotation to the teaser figure like the unit of the numerical results, higher is better or lower is better. Because some results are higher is better, e.g., PSNR while some are lower is better such as training time. - More details of the experimental setup in section 4.1 could be provided to make the implementation clearer. For instance, the authors used the LPIPS as one of the metrics but they did not specify which perceptual network is adopted since this choice may drastically affect the LPIPS score.

Questions

I have two questions: a) In the paper of HDR-NeRF, the training set and testing sets of real scenes are completely separate and have no intersection according to the implementation details. However, in the official github repository of HDR-NeRF, the training and testing sets for real experiments have intersection, which makes me very confusing. So I want to figure out, in your real experiments, did you separate the training and testing sets or just follow the official code of the HDR-NeRF? I think this is important. b) I want to know the training stability of the proposed method since I found HDR-NeRF easily collapse and need to train multiple times to make it work on some scenes.

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes, the authors have analyzed the limitations and broader impact of the method in section 4 and 5 of the supplementary pdf.

Reviewer E3vH5/10 · confidence 4/52024-07-12

Summary

This paper introduces HDR-GS, a framework designed for efficient rendering of high dynamic range (HDR) novel views. HDR-GS leverages a Dual Dynamic Range (DDR) Gaussian point cloud model that utilizes spherical harmonics for HDR color fitting and an MLP-based tone-mapper for low dynamic range (LDR) color rendering. Given an exposure time, HDR-GS can reconstruct the corresponding LDR image, achieving a form of controllable tone-mapping. The method demonstrates significant improvements over state-of-the-art NeRF-based methods in terms of both speed and image quality.

Strengths

HDR-GS achieves 1000x faster inference speed compared to HDR-NeRF. Its training is also efficient. The results shows significant improvement in image quality. The framework is sound. The authors provide a detailed derivation process, demonstrating the motivation and rationale behind the framework design. The tone mapper design enables controllable exposure time for reconstrcting LDR image.

Weaknesses

This method requires taking photos with different exposure settings at each camera position and additional HDR image data in $L_c$ to calculate the loss function. These photos and data are relatively difficult to obtain in practice. The entire pipeline is quite similar to HDR-NeRF, including the tone mapper MLP. Essentially, the authors replace the NeRF MLP with Gaussian splatting. To adapt HDR-NeRF to Gaussian splatting, the authors propose several key modifications: 1) camera recalibration and point cloud generation; 2) a constant bias $b$ in Equation 8; and 3) using $L_c$ instead of a unit exposure loss $L_u$. However, these modifications are minor and contribute only slightly. Moreover, the poor performance of the baseline model in Table 3a indicates that LDR supervision alone, without GT HDR images, is insufficient for reconstructing the HDR point cloud. This somewhat weakens the novelty. There is not ablation study for the constant bias $b$ in Equation 8.

Questions

My main concern is about the novelty issue mentioned in the Weaknesses section. Is it possible to apply HDR-GS to data where each viewpoint has a different exposure time?

Rating

5

Confidence

4

Soundness

4

Presentation

4

Contribution

2

Limitations

The authors may discuss the difficulty of data acquisition in real-world settings.

Authorsrebuttal2024-08-11

Discussion with Reviewer E3vH

Thanks for your reply. We appreciate your recognition of our work and performance. &nbsp; Actually, our HDR-GS is not simply adjusting HDR-NeRF to 3DGS. Here are some comparisons in technical details: &nbsp; (i) Tone-mapping Operation. The tone-mapping operation of HDR-NeRF converts the HDR color of the sampled 3D points to LDR color following volume rendering along every single ray. Specifically, the HDR volume rendering is formulated as $\mathbf{I}^h(\mathbf{r}) = \sum_{i=1}^{N} T_i (1 - \exp(-\rho_i \delta_i)) \mathbf{c}^h_i$, where $\rho_i$ is the volume density at the $i$-th sampled point. $\mathbf{c}_i^h$ is the HDR color at the $i$-th point. $\mathbf{I}^h(\mathbf{r})$ is the HDR color of the pixel where ray $\mathbf{r}$ lands on. The tone-mapping operation of HDR-NeRF converts $\mathbf{c}^h_i$ to the LDR color $\mathbf{c}^l_i$. 3DGS also adopts a similar point-based rendering in rasterization. So directly adapting HDR-NeRF to 3DGS should perform the tone-mapping operation following the HDR rasterization. Specifically, in Eq.(13), the point-based rendering in HDR rasterization is $\mathbf{I}^{h}(p) = \sum_{j \in \mathcal{N}} \mathbf{c}_j^h \sigma_j \prod _{k=1}^{j-1}(1-\sigma_k)$. The naive adaptation should perform tone-mapping to convert $\mathbf{c}_j^h$ to $\mathbf{c}_j^l$ along the ray landing on the pixel $p$. However, as compared in the following table, which is copied from the table in `A-2` of our rebuttal for your convenience, we found this adaptation only yields suboptimal results and speed because it can only extract limited color information on a single ray for HDR-to-LDR transformation and needs to compute many times. |Tone-mapping|Train Time (min)|Infer Speed (fps)|LDR-NE|LDR-OE|HDR| |:-|:-:|:-:|:-|:-|:-| |Ray Tracing (HDR-NeRF)|58|78|39.51|34.68|36.12| |Global Infer (Ours)|34|126|41.10|36.33|38.31| To address these issues, we design the Dual Dynamic Range (DDR) Gaussian point cloud model that performs a global tone-mapping operation converting the HDR colors of all Gaussian point clouds to LDR at one time to extract more contextual color information and accelerate the inference speed. As shown in the above table, our global tone-mapping leads to improvements of 1.59/1.65/2.19 dB and 48 fps on LDR-NE/LDR-OE/HDR and speed. &nbsp; (ii) Data Recalibration. As analyzed in Lines 174 - 181, HDR-NeRF adopts the normalized device coordinate (NDC) system that rescales the coordinates to the unit cube [-1, 1]$^3$ to help stabilize the training. Plus, the data collected by HDR-NeRF does not provide point clouds for the initialization of Gaussian point clouds. The naive and straightforward adaptation is to randomly init the positions of Gaussian point clouds within the cube [-1, 1]$^3$ and use the NDC with 2D projections to optimize. However, when we try this naive adaptation, the HDR-GS only achieves poor results of 24.45, 25.31, and 23.08 dB on HDR, LDR-OE, and LDR-NE for two reasons. Firstly, the NDC restricts the representing ability and spatial transformation of Gaussian point clouds. Secondly, training the randomly initialized Gaussian point clouds with few views (only 18) leads to an overfitting issue. To address these problems, we use the SfM algorithm to recalibrate the camera parameters and compute the initial point clouds. A naive method is to use all images to recalibrate, as mentioned in `Q-2` of our response to reviewer `tLXM`. However, this naive recalibration achieves suboptimal results of only 34.21, 36.19, and 33.75 dB on HDR, LDR-OE, and LDR-NE because the light intensity change degrades the accuracy of feature keypoints detection and matching. Hence, we use the LDR images with the same exposure time to recalibrate in Eq.(14), leading to improvements of 4.10, 4.91, and 2.58 dB on HDR, LDR-OE, and LDR-NE. &nbsp; (iii) HDR supervision. HDR-NeRF uses the ground truth CRF correction coefficient $C_0$ for HDR supervision. The naive adaptation should also use $C_0$. Yet, we found that training with $C_0$ correction is unstable and the model easily collapses. Besides, $C_0$ is unavailable in real scenes where HDR-NeRF naively sets $C_0 = 0.5$. Yet, this inaccurate $C_0$ may cause color distortion or introduce black spots, as shown in Figure 6. Thus, we use Eq.(16) as the HDR supervision for quantitative evaluation to stabilize the training process. When HDR images are not available in practice, we do not use inaccurate $C_0$ to avoid degradation. &nbsp; Besides, we resolved the data splitting issue on the real datasets, as mentioned in `Q-5` of our response to reviewer `tLXM`. &nbsp; 3DGS is a great work. Yet, directly applying 3DGS or naively adapting HDR-NeRF to 3DGS does not work well. Our work, as the first attempt, proposes an effective method to explore the potential of 3DGS for HDR imaging. We are glad to share our code, model, and recalibrated data with the community. &nbsp; Feel free to ask us if you have other questions. Looking forward to your reply.

Authorsrebuttal2024-08-13

Discussion with Reviewer E3vH

&nbsp; Dear reviewer `E3vH`, &nbsp; Thanks for your time and valuable comments. We appreciate your recognition of our work and effort. Could you please let us know if our response addressed your concerns about the comparison between the simple adaptation of HDR-NeRF for 3DGS and our proposed HDR-GS? We sincerely appreciate your willingness to raise the score. Just a friendly reminder that the scores have not been updated in the openreview system yet. Please feel free to ask us if you have any other questions. We are looking forward to your reply. &nbsp; Best, Authors

Reviewer E3vH2024-08-13

After reading the authors' responses, most of my concerns have been addressed, so I have raised the score.

Reviewer E3vH2024-08-10

Thank you to the authors for their response; it addressed my concerns regarding HDR supervised training. As for the comparison with HDR-NeRF, I believe most of the differences arise from the necessary adjustments when transitioning between the representation and rendering pipeline of NeRF and 3DGS. These adjustments seem to be natural solutions and do not represent significant novelty. The improvements in rendering quality and speed of HDR-GS also stem from the inherent characteristics of 3DGS. Could the authors point out any technical modifications that are not simply required for adapting the NeRF representation from HDR-NeRF to 3DGS, but that also enhance performance? Or could they explain why some adjustments made during the transition, which could have been done more simply, are superior in their current form? In any case, I appreciate the effort the authors put into this work, which has improved the performance of the results. Considering the rebuttal that has addressed some of my concerns, I am willing to slightly raise the review score.

Reviewer VwS82024-08-13

I have read all the reviews and the authors' responses. Most of my concerns have been addressed and I am inclined to keep my previous rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC