IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera

1. The setup of the real-world experiments lacks validity. Specifically, for the evaluation on the TUM-VIE dataset, the qualitative results alone are insufficient. The authors should also include quantitative analysis using no-reference image quality assessment metrics to provide a more comprehensive evaluation.

2. Although the proposed method can operate using an event stream as input to reconstruct 3D Gaussians, it still relies on uniform event stream as input. The proposed method is, therefore, limited by the density of event data streams, which restricts its practical applications.

3. Despite the detailed comparison of the quality of rendered images, the efficiency of the training and rendering process is not included, which is an important metric of NVS methods. Extra comparisons with other methods on training time and inference FPS would help better evaluate the proposed method.

4. This method is valuable for addressing event-based visual odometry. However, the authors focus more on the NVS task, and using Gaussian functions to reconstruct grayscale scenes seems less relevant, as they are mainly suited for head-mounted devices, which reduces the method’s rationale.

Beyond this I have mainly minor comments and nitpicks:

l.117, the sentence contains a grammatical error and should be revised. Specifically, "IncEventGS conduct..." should be corrected to "IncEventGS conducts...".

l.142, the expression should be standardized by changing "se3" to "se(3)" for clarity and consistency.

l.162~186, I think the re-initialization process is vital to the method, but the main figure of the pipeline does not reflect this which may generate some confusion with readers not familiar with the method.

1. The motivation of this paper is weak, as the paper claims, "Due to its unique asynchronous and irregular data capturing process, limited work has been proposed to apply neural representation or 3D Gaussian splatting for an event camera". I think the authors should discover the reasons behind it rather than the superficial phenomenon.

2. The title is "Pose-free." Why the author did this is not explained. I think that although no pose ground truth is provided, using conventional slam pipelines actually provides this variable implicitly. Conventional slam pipelines will be more robust than pose estimators, which use deep learning methods.

3. This paper mentions several times “due to its time continuous, sparse, asynchronous and irregular data capturing characteristics.” I don't think the authors have solved this problem; they are still taking the approach of stacking events into the frame.

4. In line 62, citation duplication.

5. The contribution needs to be rewritten, which is just like changing the representation from Nerf to GS. However, this work has already been done.

6. "The main insight of IncEventGS is to accumulate incoming event data into chunks and treat each chunk as a special "image". This is not a contribution and does not need to be emphasized.

7. In line 216 and 307, C in (3) and equation 6.

The manuscript is clearly written but does not explain in a precise and in-depth manner how it is carried out. In other words, the concepts expressed are only shown at a high level, without delving into small key details, such as the “continuous time trajectory parameterization” or how “the camera poses (T_k) and (T_{k+\Delta t}) can be interpolated,” and how exactly to “render two grayscale images (i.e., (\hat{I}k) and (\hat{I}{k+\Delta t})) from the previously recovered 3D-GS,” which makes it very difficult to reproduce the results.

Although the manuscript mentions some studies related to 3D-GS and event cameras, it does not mention 3D-GS works that perform 3D reconstruction or novel view synthesis with pose-free cameras and frame-based cameras.

1. While I acknowledge that this paper is the first to explore 3D reconstruction using a single event camera combined with 3D ground segmentation without camera poses, its novelty appears to be limited. There are existing works using traditional RGB cameras for 3D reconstruction without relying on camera poses, and the approach of directly accumulating events into event images does not clearly highlight significant contributions to the field, whether from the image-based 3D ground segmentation community or the event-based community. I encourage the authors to articulate the specific technical contributions of this work.

2. I recommend that the authors include more examples of extreme scenarios, such as high-speed motion and low-light conditions, alongside comparisons with RGB images. This could better demonstrate the advantages of using event cameras for 3D reconstruction.

3. Regarding the possibility of achieving colored 3D reconstruction, can this method be applied? Since there are existing color event cameras, could the authors obtain data from such cameras to create an example of colored reconstruction?

4. The writing could be further improved in several ways: a) The title in line 97 should be bolded and capitalized. b) Section 3.2 does not require an extensive explanation of event camera principles and image accumulation. c) The font sizes in Tables 1 and 2 should be made consistent.