We sincerely thank the reviewer for the time with our paper. While we have provided detailed responses to address the main concerns below, one thing needs to be highlighted first: we respectfully cannot agree that our work "lacks innovation" due to inheriting certain “existing modules”. In contrast, the innovation of integrating NLOS and Mamba is a consensus of all other reviewers and has been well recognized.

Q1: Motivation for Mamba.
Reply: The motivation are twofold:

1. Causality in Transient Measurement: Our method starts by considering the causality inherent in transient measurement. To capture this long-ranging causal information, we employ Mamba. As shown in Tab. 1 (IDs 0 and 4) of the rebuttal 'global' PDF, disabling the causal operation in our method results in a significant drop in performance metrics, demonstrating the effectiveness of our approach in exploring causality.

2. Efficiency and Scalability of Mamba: Mamba has rapidly become popular in various tasks due to its linear computational complexity and global modelling capability. For comparison with VIT, we replace the ST-Mamba and Cross ST-Mamba with the plain attention block and the cross attention block. Note that the number of the core operators is the same. As shown in Fig.2 of the rebuttal 'global' PDF, our method exhibits exceptional linear scaling performance. As the input resolution increases, our method will use significantly less GPU memory.

Q2: Lack innovation.
Reply: There are two points to address:

1. The challenge of using Mamba for NLOS: The Mamba architecture, while effective for 1D sequential data, faces challenges in NLOS imaging due to its unidirectional processing mode. The 3D NLOS transient measurements are inherently partially causal and high-dimensional. Specifically, the histogram of each scanning point along the temporal axis exhibits causality, but the scanning points themselves along the spatial axis are non-causal. This non-causality and high dimensionality hinder Mamba from effectively capturing the underlying features. To overcome these limitations, we propose ST-Mamba and Cross ST-Mamba mechanisms, which are specifically designed to exploit and integrate the deep features within under-scanning transient measurements for NLOS imaging tasks.

2. The development of NLOS field: As we know, one advantage of the AI community is gradually building on top of existing works, which eventually leads to both solid and efficient solutions for many practical applications. When we look back, there are a number of works in this community with seemingly incremental components but have made a great impact on future research. Being confident in our results, we are ready to release the source code to the public. We would appreciate the opportunity to present this work as the new state-of-the-art together with its repository at this conference, which allows other researchers in this scientific field to build on top of it and improve the field further with an enhanced baseline.

Q3: Detailed description for Sec. 4.2, 4.3.
Reply: In Sec 4.2, we have provided a detailed explanation of the data flow, dimension change, and operator, accompanied by the schematic diagram Fig.4 for understanding. The core computational operators (SSMs) adhere to the original design, and thus we did not elaborate on them with additional formulas, as it may be redundant.

Regarding the hidden volume reconstruction, the feature extraction module comprises three 3D residual blocks for extracting shallow features and downsampling along the temporal axis. The volume refinement module is composed of three 3D convolutions and three interlaced 3D residual blocks. Each residual block comprises two 3D convolutions followed by a ReLU activation and a residual connection. We will include the explanation in Sec 4.3. Moreover, we will revise the colloquial words, e.g., 'enter', 'pure', 'huge', etc., ensuring clarity and professionalism in our presentation.

Thanks for bringing up these points, and we will revise the corresponding components.

Q4: Ablations for ST-Mamba and Cross ST-Mamba.
Reply: As suggested, we have conducted more ablation studies on ST-Mamba and Cross ST-Mamba. The quantitative results are listed in Tab.1 of the rebuttal 'global' PDF.
We individually replaced the ST-Mamba (blue blocks in Fig.2 of the manuscript) and Cross ST-Mamba (orange blocks in Fig.3 of the manuscript) with a plain attention block and a cross attention block. By comparing IDs 1, 3, and 4, it is evident that our method performs best when both ST-Mamba and Cross ST-Mamba are employed. Additionally, we excluded only the Cross ST-Mamba while maintaining the same number of SSMs as in our final method (ID 2). This further demonstrates the effectiveness of the proposed Cross ST-Mamba.
Thank you for your valuable suggestion. We will include the additional ablation studies in the revised manuscript.

Q5: Ablations for spatial and temporal Mamba.
Reply: To thoroughly verify the impact of spatial Mamba (S-Mamba) and temporal Mamba (T-Mamba), we have conducted additional comprehensive ablation studies. In these studies, we replaced the spatial Mamba with temporal Mamba and vice versa to validate their individual contributions to the network. The results are listed as follows:

It can be seen that spatial and temporal Mamba contribute differently to the reconstruction performance, and fusing both of them boosts the overall performance.

We truly appreciate your suggestions and will include the detailed settings and descriptions in the revised manuscript.

We are highly encouraged by the positive recommendations and comments from the reviewer. The main concerns are addressed below.

Q1: Lack video results.
Reply: We sincerely appreciate your kind reminder. After preparing the video as requested, however, we find that external links are not allowed in rebuttal, even with an anonymous one. We will include the qualitative video results in the final supplementary materials.

Q2: Motivation for using SSMs to enforce spatio-temporal consistency.
Reply: There are two points to address:

1. By employing a well-designed structure to fully exploit the causality of transient data, our single-frame method (without cross ST-Mamba) outperforms the existing methods. Our experimental results demonstrate that SSMs have significant application potential for transient data with partial causality. Therefore, we aim to use SSMs to further maintain spatio-temporal consistency.

2. Additionally, we conduct an experiment using cross attention to maintain spatio-temporal consistency, with results listed in Tab. 1 (ID 3) of the rebuttal ‘global’ PDF. It can be seen that using Cross ST-Mamba (ID 4) yields the best results. This does not imply that cross-attention mechanism is unsuitable for maintaining spatio-temporal consistency; rather, it indicates that using Cross ST-Mamba is more effective for transient data with partial causality.

Q3: Contribution in cross ST-Mamba component.
Reply: As shown in Tab.1 of the original manuscript, we provide the results of the single-frame version of our method, Ours-S. Note that Ours-S excludes the cross ST-Mamba but possesses the same number of SSMs as Ours. The multi-frame model (Ours) surpasses the single-frame model across all metrics by a large margin, which showcases the strength of the proposed cross ST-Mamba in integrating information from multiple transient frames.

We are highly encouraged by the positive recommendations and comments from the reviewer. The main concerns are addressed below.

Q1: The explanation for the causality in transients.
Reply: First, let us review the definition of causal signals: Causal signals refer to signals where the output at any given time depends not only on the current input but also on past inputs. These signals exhibit a cause-and-effect relationship where the effect (output) is directly influenced by prior causes (inputs). In other words, the response of a system or process at any time is a function of its history.
Next, let us apply the definition to transient measurement and explain causality. For each sampling point, the temporal histogram indeed contains all the geometry information of the hidden object because the spherical waves are scattered back by the illuminated object. Considering any point on the temporal axis of the histogram, the value at this point is only related to a target point on the hidden object and any point closer to the relay wall than the target point (shorter flight). Clearly, this conforms to the definition of causality. It is also worth mentioning that the histogram is obtained through periodic cumulative accumulation, which is a practical operation to compromise the dead time of SPAD. This does not affect the causal nature.
We will include this further explanation in the manuscript.

Q2: The operations on the temporal dimension.
Reply: Of course, the transient data can be processed while always maintaining the temporal dimension. However, performing convolution operations, dot product operations, etc., on three-dimensional data consumes a significant amount of computational resources. Therefore, it is NOT a common practice not design networks directly and continuously at high dimensions.
To some extent, we agree that The temporal information in low spatial-resolution transients might be compromised by down-sampling and up-sampling operations. As discussed above, this is the trade-off between performance and consumption. We emphasize this point to highlight that the reconstruction of the temporal dimension is crucial in transient measurement recovery tasks, as validated by the [1].

[1] Jinye Miao, Enlai Guo, Yingjie Shi, Fuyao Cai, Lianfa Bai, and Jing Han. Super-resolution non-line-of-sight imaging based on temporal encoding. Optics Express, 31(24):40235–40248, 2023.

Q3: Comparison of the SR module separately with USM.
Reply: As suggested, we compare the capabilities of the transient super-resolution networks of USM[1] and our method. To obtain quantitative results, we first generate the recovered high-resolution transient measurements and then conduct the reconstruction using the traditional method RSD[1], instead of the deep neural network. For USM, the metrics PSNR/SSIM/RMSE/MAD are 19.25/18.93/0.2206/0.1000. For our method, the metrics PSNR/SSIM/RMSE/MAD are 21.48/17.08/0.2120/0.0917. Except for SSIM, our method surpasses USM. The decrease in SSIM may originate from the artifacts interference.
To provide a more comprehensive evaluation, we introduce another metric ACC, from[2], which indicates the foreground recovery degree. The ACC for USM is 10.49, while for our method, it is 31.22, demonstrating that our method generates cleaner and higher-quality transient measurements. Qualitative results are provided in Fig.1 of the rebuttal 'global' PDF, further demonstrating our superiority comprehensively.

[1] Xiaochun Liu, Ibón Guillén, Marco La Manna, Ji Hyun Nam, Syed Azer Reza, Toan Huu Le, Adrian Jarabo, Diego Gutierrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572(7771):620–623, 2019.
[2] Jun-Tian Ye, Xin Huang, Zheng-Ping Li, and Feihu Xu. Compressed sensing for active non-line-of-sight imaging. Optics Express, 29(2):1749–1763, 2021.

We are highly encouraged by the positive recommendations and comments from the reviewer. The main concerns are addressed below.

Q1: Suggestions for the title.
Reply: We appreciate your insight that renaming it to "NLOS Imaging by Using Dynamic Cues" would more accurately reflect our approach and its capabilities. We will take this into consideration in our revisions to improve clarity and accurately represent our work.
Thanks for your valuable feedback.

Q2: Discussion about the prior art on dynamic NLOS.
Reply: Thanks for your insightful comments and the suggested references. ''Real-time Non-line-of-Sight Imaging of Dynamic Scenes.'' is based on a non-confocal setup for dynamic imaging, which has been discussed in Sec.2. ''Keyhole Imaging: Non-Line-of-Sight Imaging and Tracking of Moving Objects Along a Single Optical Path'' is the technology using a fixed scanning and sampling point for dynamic imaging, and it indeed represents a third type of prior work in Sec.2.

We will include ''keyhole imaging'' as a third type and cite additional relevant papers for a more comprehensive discussion. Thanks again for your valuable suggestions. We believe these additions will significantly strengthen our manuscript and clarify the contributions of our work.

Q3: A preliminary section for Mamba.
Reply: Thanks for your suggestion, we will include a recap as follows:

State Space Model (SSM)

The State Space Model (SSM) is employed to describe the linear time-invariant systems. The system processes the 1D input sequence $x(t)\in \mathbb{R}$ by propagating them through the intermediate hidden states $h(t)\in \mathbb{R}^{N}$, ultimately generating output sequences $y(t)\in \mathbb{R}$. Typically, SSM can be expressed as the linear ordinary differential equation:

$$h^{'}(t)=\mathbf{A} h(t) + \mathbf{B}x(t),\quad y(t) = \mathbf{C}h(t) \quad \quad \quad1.$$

where $A\in \mathbb{R}^{N\times N}$ is the state matrix, $B,C\in \mathbb{R}^{N\times 1}$ are the projection parameters. After discretization via the timescale parameter $\Delta$[1], Eq.1 is formulated as:

$$h_{t} =\mathbf{\bar{A}}h_{t-1} + \mathbf{\bar{B}}x_t,\quad y_t = \mathbf{C}h_t \quad \quad \quad \quad \quad \quad \quad 2.$$

where $\mathbf{\bar{A}}=\text{exp}(\Delta \mathbf{A})$, $\mathbf{\bar{B}}=(\Delta \mathbf{A})^{-1}(\text{exp}(\Delta\mathbf{A}-I)\cdot \Delta\mathbf{B}$. Besides, Eq.2 can be transformed into a convolutional operation which is suitable for hardware. Recently, Mamba[1] employes data-dependent mechanisms, including the learnable parameters $B$,$C$, and $\Delta$, as well as parallel scanning. Mamba has rapidly become popular in various tasks[2,3,4,5,6] due to its linear computational complexity and global modelling capability.

[1] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021.
[2] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.
[3] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024.
[4] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Visionmamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024.
[5] Zifu Wan, Yuhao Wang, Silong Yong, Pingping Zhang, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Sigma: Siamese mamba network for multi-modal semantic segmentation. arXiv preprint arXiv:2404.04256,2024.
[6] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739,2024.

Q4: Another comparison with more sampling points.
Reply: Thanks for your thoughtful suggestion. However, we did not explore this specific setup due to the following considerations:

1. In a confocal setup, there is a trade-off between the number of sampling points and the exposure time for each point, regarding imaging quality. Our current setup achieves 4 FPS with an 800 µs exposure time and a 16x16 scanning grid. Increasing the number of scanning points to maintain an equal amount of data would reduce the frame rate to 1.3 FPS, which compromises our goal of high-speed imaging.
2. Keeping the total number constant, or increasing the scanning-grid would be unfair to multi-frame algorithms, as their scanning-grid remains unchanged.
3. It is important to note that the amount of training data used for the single-frame methods is consistent with that used for our final method, ensuring the fairness of the comparison for data-driven approaches.

Thanks for your understanding.

Thank you for your thoughtful review and we are pleased that our response effectively addressed your concerns.

As suggested, we will include detailed descriptions of the imaging rate and the relevant scenes, further highlighting the significance of using a confocal setup for dynamic imaging.

Thanks again for the time you've invested in helping us improve our work.