Learning Differential Pyramid Representation for Tone Mapping

1. There are textual overlaps with [1] and [2]. Lines 186-199 in this submission are identical to Section 3.1 in [1] except for notations. Line 287-288 is identical to the caption of Figure 4 in [2]. Lines 363-366 are nearly identical to Section 3.2 in [2]. All of them are without proper citation.

2. The Learnable Differential Pyramid (LDP) module and Iterative Detail Enhancement (IDE) module raise concerns, and please refer to the Questions Section.

3. The numeric citations do not follow the formatting requirements.

4. Some minor errors: The circled $C_c$ notation in Figure 2 differs from the circled $C$ in Equal 4. The lines in Table 4 are misaligned. Reconfirm the usage of uppercase of "W/O" in Figure 5.

[1] He, Jingwen, et al. "Conditional sequential modulation for efficient global image retouching." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16. Springer International Publishing, 2020.

[2] Zhang, Feng, et al. "Lookup table meets local laplacian filter: pyramid reconstruction network for tone mapping." Advances in Neural Information Processing Systems 36 (2024).

1. Pyramid decomposition has been applied in [23][4]. For the low-resolution enhancement part, the method proposed in this paper is like a combination of CSRNet and 3DLUT, which seems to have limited innovation. In addition, this paper does not give full play to the speed advantage of pyramid decomposition.
2. The processing speed of the proposed method is slow, which is nearly 40 times slower than the LUT method.

1. The reviewer can see that the paper lacks of literature review. For example, for tone mapping

• Unsupervised HDR Image and Video Tone Mapping via Contrastive Learning
• Fast global tone mapping for high dynamic range compression
• A Real-Time Semi-Supervised Deep Tone Mapping Network
• A Low-Latency Noise-Aware Tone Mapping Operator for Hardware Implementation with a Locally Weighted Guided Filter

2. Model

• The design and the explanation of the model is confusing. The reviewer find that the explanation does not match the figure.
• It is not clear which one is the main contribution in the model, the reviewer believe that the LDP is the main, others are employed from other papers.

3. The paper lacks of a strong motivation of the paper. The reviewer understand that the target is the LDR display, but did not see that when reading the paper.

4. The reviewer understand that the paper follow [4] when comparing with other methods. However, the reviewer think that comparing a ToneMapping method with image enhancement methods is not fair. If the author did that, were they retrained with the dataset that used in the proposed method?

Computational Complexity: While effective, DPRNet is computationally intensive, particularly with high-resolution images. The paper’s discussion of runtime and computational load is limited, which might impact its scalability for real-time or mobile applications.

Limited Benchmark Comparison: Although the authors provide comparisons with recent state-of-the-art methods, some relevant techniques are omitted, particularly those with similar high-frequency-focused approaches (e.g., S. Moran et al., ‘‘CURL: Neural curve layers for global image enhancement,’’ in ICPR, 2021 and L. Zhao et al., "Learning Tone Curves for Local Image Enhancement," in IEEE Access, 2022).

Generalization on Noisy Datasets: DPRNet struggles on datasets like MIT FiveK due to noise and overexposure, as noted by the authors. The paper could benefit from discussing potential solutions or additional noise-handling components.

User Study: While a user study is included, the scope could be broader to include users with different levels of experience in HDR processing, providing a more diverse evaluation of perceived quality.