Lookup Table meets Local Laplacian Filter: Pyramid Reconstruction Network for Tone Mapping

Tone mapping aims to convert high dynamic range (HDR) images to low dynamic range (LDR) representations, a critical task in the camera imaging pipeline. In recent years, 3-Dimensional LookUp Table (3D LUT) based methods have gained attention due to their ability to strike a favorable balance between enhancement performance and computational efficiency. However, these methods often fail to deliver satisfactory results in local areas since the look-up table is a global operator for tone mapping, which works based on pixel values and fails to incorporate crucial local information. To this end, this paper aims to address this issue by exploring a novel strategy that integrates global and local operators by utilizing closed-form Laplacian pyramid decomposition and reconstruction. Specifically, we employ image-adaptive 3D LUTs to manipulate the tone in the low-frequency image by leveraging the specific characteristics of the frequency information. Furthermore, we utilize local Laplacian filters to refine the edge details in the high-frequency components in an adaptive manner. Local Laplacian filters are widely used to preserve edge details in photographs, but their conventional usage involves manual tuning and fixed implementation within camera imaging pipelines or photo editing tools. We propose to learn parameter value maps progressively for local Laplacian filters from annotated data using a lightweight network. Our model achieves simultaneous global tone manipulation and local edge detail preservation in an end-to-end manner. Extensive experimental results on two benchmark datasets demonstrate that the proposed method performs favorably against state-of-the-art methods.

Paper

References (40)

Scroll for more · 28 remaining

Similar papers

Peer review

Reviewer h55c6/10 · confidence 4/52023-06-20

Summary

In this paper, the authors propose an effective end-to-end framework for HDR image tone mapping tasks, combining global and local enhancements. The proposed framework utilizes the Laplacian pyramid decomposition technique to handle high-resolution HDR images effectively. This approach significantly reduces computational complexity while simultaneously ensuring uncompromised enhancement performance. Global tone manipulation is performed on the low-frequency image using 3D LUTs. An image-adaptive learnable local Laplacian filter is proposed to progressively refine the high-frequency components, preserving local edge details and reconstructing the pyramids. Extensive experimental results on two publically available benchmark datasets show that our model performs favorably against state-of-the-art methods for both 480p and 4K resolutions. The major contribution of this work is combination of high frequency lap filter and low frequency global tone mapping for acheiving new SOTA results for tone mapping.

Strengths

New SOTA results for tone mapping using local learnable lap filters for refining high frequency details.

Weaknesses

Generalization to other datasets may need retraining.

Questions

How about mid frequency treatment ?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Improvement may be limited to scenes with complex lightings.

Reviewer zrx67/10 · confidence 5/52023-07-02

Summary

The paper introduces a novel framework for converting high dynamic range (HDR) images to low dynamic range (LDR) images, a crucial task in digital imaging. The proposed framework aims to improve the balance between image enhancement and computational efficiency by incorporating global and local tone mapping operators. Traditional methods are either global, processing the whole image uniformly but often missing fine details, or local, capturing fine details with higher computational requirements. To overcome this dichotomy, the authors propose integrating 3D LookUp Tables (3D LUT) as a global operator and local Laplacian filters as a local operator into one end-to-end system. The 3D LUTs are used for global tone manipulation, while local Laplacian filters refine high-frequency edge details. Moreover, the authors introduce a learnable local Laplacian filter (LLF) and a lightweight network that progressively learns parameter value maps from annotated data for the local Laplacian filters. This approach reduces manual tuning, which is a limitation of existing methods. The results of the extensive experiments on two benchmark datasets demonstrate that the proposed framework outperforms state-of-the-art methods in terms of both image quality and computational efficiency. The paper contributes to digital imaging by offering an efficient solution to HDR image tone mapping while preserving important details.

Strengths

The paper is well-written and structured. The authors clearly explain the problems with existing methods and why their proposed framework offers a solution. Concepts are clearly defined, and the methodology is explained in a detailed and straightforward manner. Diagrams, equations, and examples effectively support the reader's understanding.

Weaknesses

- The paper misses the discussions and comparisons of many related methods [1-3]. These methods are all published in good venues within the past five years and thus deserve a fair discussion. - Why the evaluation metrics do not contain TMQI, Fidelity, and Naturalness? These three evaluation metrics are common practice in TMO papers (see [1] for example). As tone mapping is a subjective task, using PSNR or SSIM to evaluate does not provide much information. - Some of the compared methods, such as CLUT [39], are tacking low-light image enhancement. Why did the authors compare these methods? The input of tone-mapping is an HDR radiance map, while the input to low-light image enhancement is an already tone-mapped non-linear image. The comparison is unfair and could mislead future work. - I strongly suggest the authors revisit the goal of their proposed method. Is it a TMO (tone mapping operator) or low-light image enhancement (LLIE)? I really don't see many TMO methods in the comparison section. If that is the case, I would suggest the authors rename the paper title to low-light image enhancement and change all the terms in the paper. - If the proposed method indeed is a TMO, then the first step should be LDR to HDR as the input is a low-light image. Then the authors should also discuss the related work [4-6] and design choices for LDR2HDR. - While the authors tested their framework on two benchmark datasets, it would have been beneficial to see the system evaluated on more diverse datasets, including real-world photographs, video frames, and even synthetic images. This could have offered a more comprehensive understanding of the method's robustness and generalizability. - The paper lacks a complex analysis of the proposed framework. A comparison of the computational complexity and memory usage between the proposed model and existing methods would give a clearer picture of the performance trade-off. - The authors have not thoroughly discussed the potential limitations of their proposed method. Understanding what the model struggles with or where it might fail is essential for future improvements and applications. [1] Liang, Zhetong, et al. "A hybrid l1-l0 layer decomposition model for tone mapping." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. [2] Su, Chien-Chuan, et al. "Explorable tone mapping operators." 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2021. [3] Vinker, Yael, Inbar Huberman-Spiegelglas, and Raanan Fattal. "Unpaired learning for high dynamic range image tone mapping." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021. [4] Liu, Yu-Lun, et al. "Single-image HDR reconstruction by learning to reverse the camera pipeline." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020. [5] Santos, Marcel Santana, Tsang Ing Ren, and Nima Khademi Kalantari. "Single image HDR reconstruction using a CNN with masked features and perceptual loss." arXiv preprint arXiv:2005.07335 (2020). [6] Chen, Xiangyu, et al. "Hdrunet: Single image hdr reconstruction with denoising and dequantization." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021.

Questions

See the section above.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

The paper does not discuss the limitation of the proposed method. I believe the proposed method will not have a potential negative societal impact.

Reviewer 1rH87/10 · confidence 5/52023-07-06

Summary

This paper introduces an effective end-to-end framework for HDR image tone mapping. The network performs both global tone manipulation and local edge details preservation within the same model. The proposed method learns an image-adaptive learnable local Laplacian filter for local edge details preservation.

Strengths

1. This paper is well-organized and easy to read. 2. The experiment results show its effectiveness on two benchmarks, as well as the visual results. 3. The proposed learnable local Laplacian filter shows remarkable improvement when integrated with image-adaptive 3D LUT.

Weaknesses

1. Only limited benchmarks are valued, unlike previous work like CLUT, evaluated on both FiveK, HDR+, and PPR10K, so more evaluation results are encouraged. 2. There is no inference time presented in this work, which is another key information for enhancement applications.

Questions

1. Why are the key metrics of baselines reported in this paper significantly lower than those reported in their original paper? For example, the PSNR of CLUT[39] on the HDR+ dataset is 26.98dB, while this figure is only 26.05dB in your Table 1, the PSNR of sLUT [36] on the HDR+ dataset is 26.94dB, while this figure is 26.13dB in your paper, all these figures are much higher than yours. The same issues occur in Table 2.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The limitation is discussed in Supp materials.

Reviewer g4yF3/10 · confidence 4/52023-07-07

Summary

The paper proposes an algorithm for the HDR image tone mapping task. It is an end-to-end framework combining global and local enhancements. It uses the Laplacian pyramid decomposition technique. It achieves a balance between computational complexity and performance. Global tone manipulation is performed on low-frequency component. Experiment shows that it out-performs SOTA.

Strengths

It achieves a balance between computational complexity and performance. Experiment shows that it out-performs SOTA. It is intuitive to do global tone manipulation on low-frequency image, while using image-adaptive learnable local Laplacian filter for high-frequency component. Nice ablation studies are performed to quantify the improvements from different modules of the whole algorithm.

Weaknesses

I wonder why the the transformer weight predictor and the basis 3D LUTs fusion block only exist in the coarsest level of the pyramid. Is there any way to integrate them into all levels? If integrate them into all levels, how much additional computational burden will be introduced? The quantitative comparison on HDR+ dataset shows that CSRNet and LUT are both smaller than the proposed algorithm. Is there any chance to compare with them with similar amount of computation? E.g. if the proposed algorithm is reduced to 37K in terms of number of parameters, will it still outperform CSRNet? LUT will use heavy storage. How to compare the proposed algorithm with SOTA in terms of data storage? Number of parameters alone will not tell the whole picture of the efficiency, what is the real running time comparison?

Questions

see weaknesses

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

did not talk about potential negative societal impact

Reviewer nHJa6/10 · confidence 4/52023-07-08

Summary

This paper proposed a pyramid reconstruction network for tone mapping via a combination of existing lookup table and local Laplacian filter techniques, by utilizing closed-form Laplacian pyramid decomposition and reconstruction. Specifically, the image-adaptive 3D LUTs is employed to manipulate the tone in the low-frequency image by leveraging the specific characteristics of the frequency information. The local Laplacian filter is utilized to refine the edge details in the high-frequency components in an adaptive manner. Extensive experimental results on two benchmark datasets demonstrate that the proposed method performs favorably against state-of-the-art methods.

Strengths

- Originality: The idea of optimize existing Laplacian pyramid decomposition and reconstruction based tone mapping method via the combination of lookup table and local Laplacian filter techniques is interesting and reasonable. Besides, the proposed image-adaptive learnable local Laplacian filter can preserve the local edge details. - Clarity: This paper is well-organized and easy to follow.

Weaknesses

- The algorithm description is not clear enough:\textit{e.g.,} the definition of objective function for the image-adaptive learnable local Laplacian filter in Eq. (4) is missing. - Insufficient experimental analysis: - The runtime analyse are not provided. - Quantitative comparison of different methods on HDR+ datasets is missing.

Questions

- How about the performance for the setting with image-adaptive Local Laplacian Filter disabled? This result helps to determine the effectiveness of the proposed technology. - The visual comparison in Fig. 4 is not easy to understand. What does the heat map in the upper left corner mean?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Not applicable

Reviewer 1rH82023-08-14

A raise of the final rating from 6 to 7.

My concerns were well addressed and I would love to see this manuscript on the accepted list.

Reviewer nHJa2023-08-19

Thank you for the rebuttal. All my concerns were well addressed.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC