Edge-Sampler: Efficient Importance Sampling for Neural Implicit Surfaces Reconstruction

-It seems that the ablation study of the proposed method is not sufficient. Only different fitting strategies are shown in Figure 6. More ablation studies related to other strategies are needed. -It seems that there is no captions for Figure 4, and there is also not any explanations in the main part of the paper for this figure.

• Novelty: efficient sampling has been studied for relatively a long time since NeRF was proposed. From my perspective, it is not exciting to see research on such a topic. I would expect a significant improvement in efficiency/quality if the paper is to be accepted by a top-tier conference.

• Sampling time comparison: I cannot get the key points from Table 1. Which is the proposed method (Edge32 w/ GP)? Looks like the original NeuS sampling is already the best.

• Comparison with NGP: from Table 3 it looks like NGP is already fast and with high quality. This is also my major concern, different representations (MLP, NGP, K-planes, 3D Gaussians) would have very different sampling strategies, making the method hard or even impossible (e.g. 3D Gaussians) to apply to those representations. Also, powerful new representations would weaken the contribution (efficiency) of the paper.

• The authors only conduct experiments on 4 DTU scenes, which is not convincing to me.

• Too many equations in the paper. I would further simplify equations 1-8 as it is only a recap of previous works.

• The writing of the paper could be further improved.

The method description shows many details. It is better to attach an algorithm to summarize the entire sampling mechanism.

The experiments need to demonstrate the advantage of the proposed four-step edge sampler clearly.

• The model configurations of the compared methods in section 5.1 need to be clarified.
• The DTU dataset comprises many scenes, and it is better to include the complete results for performance comparison.
• It lacks a table to compare the reducing sample counts against other methods.
• Figure 4 has no caption.
• What is the unit of time in Tables 2 and 3? Is it fair to compare time using the batch size more than one?
• The visualization differences in Figures 4 and 6 are hard to perceive.
• What causes the performance degradation of NeuS in Table 3 compared to their paper?

• Theoretical limitation

• The proposed approach assumes that narrower sampling leads to more efficient surface reconstruction. However, this assumption doesn't hold when the optimization process begins with an unknown underlying surface. This might result in slower convergence, particularly when the initialization is far from optimal. The manuscript does not currently acknowledge or address this limitation.

• Incomplete manuscript

• The evaluation of reconstruction accuracy is based on only four scenes from the DTU dataset. The selected results are only for Lambertian surfaces, while many prior works showcase results of specular surfaces of DTU dataset as well. Without insights into the results for the entire dataset, it's challenging to assess the performance comprehensively.

• The choice of a single baseline (NeuS) for surface accuracy evaluation is limiting, ignoring the recent developments in surface reconstruction techniques, including but not limited to [Yariv et al, 2021] [Wang et al, 2022] [Fu et al, 2022] [Wang et al, 2023] [Li et al, 2023] [Li et al, 2023] [Li et al, 2023] and etc.

• While the manuscript claims improved efficiency, the speed comparison neglects the Occupancy Grid technique proposed in Instant NGP [Muller et al, 2022]. Furthermore, the proposed approach is slower than NeuS unless NeuS' interpolation technique is employed. Moreover, an important baseline is missing, which is the combination of the proposed network and NeuS.

• Figure 4 doesn’t have captions.

• Table 3 time doesn’t have units.

• Mismatch motivation and implementation

• The manuscript contains a significant discrepancy between its stated motivation and its actual implementation. In the introduction, the manuscript emphasizes the advantages of the Laplace distribution as seen in VolSDF [Yariv et al, 2021] compared to other SDF-based rendering methods. However, in the methodology section, the proposed sampling approach appears to be predominantly focused on enhancing VolSDF. Surprisingly, in all the experiments, it appears that the approach utilizes NeuS' formulation instead. It is important to justify if this is the case and why so.

References:

Yariv, L., Gu, J., Kasten, Y., & Lipman, Y. (2021). Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34, 4805-4815.

Wang, Y., Skorokhodov, I., & Wonka, P. (2022). Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35, 1966-1978.

Müller, T., Evans, A., Schied, C., & Keller, A. (2022). Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4), 1-15.

Fu, Q., Xu, Q., Ong, Y. S., & Tao, W. (2022). Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35, 3403-3416.

Wang, Y., Han, Q., Habermann, M., Daniilidis, K., Theobalt, C., & Liu, L. (2023). Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 3295-3306).

Li, Z., Müller, T., Evans, A., Taylor, R. H., Unberath, M., Liu, M. Y., & Lin, C. H. (2023). Neuralangelo: High-Fidelity Neural Surface Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 8456-8465).