Exploring Source View Capability: Improve Generalizable 3D Reconstruction with Multi-view Context from Source Views

• The transition from Equation (6) to (7) appears incorrect. For example, the second term of Equation (5) is removed in Equation (6) but reappears in Equation (7). Additionally, the square of the sum in the first term becomes the sum of squares.
• When $\alpha$ in Equation (7) is set to 1.0 to remove the last term, it is unclear how this the result loss function $\mathcal{L}_{\rm sgc}=(\sum q_i)^2$ maintains the constraint of having at most two adjacent non-zero samples, although it holds for Equation (6).
• Improvements appear limited, as shown in Table 1, with gaps of only 0.62, 0.004, and 0.004 in terms of PSNR/SSIM/LPIPS on the NeRF synthetic dataset. Similar trends are observed on the LLFF dataset and on the comparisons to pixelSplat in Table 2.
• Qualitatively, as shown in Figure 4, the proposed constraint provides minimal improvement; only one scene (second column) out of six shows clear improvement, while the rest exhibit marginal gains. A similar pattern appears in Figure 6 for 3DGS baselines.
• Figure 5 seems to be a cherry-picked example, as it shows the most substantial improvement from Figure 4. Additional examples would better demonstrate the method's benefits.
• The paper presents itself as a 3D reconstruction method but only includes novel view synthesis results, lacking explicit 3D reconstruction results.

• Limitations of the method. In the experiments, regularization is applied based on the source view's attention map along the target ray/epipolar line. It appears that the proposed depth regularization may be directly compatible only with transformer-based generalizable NeRFs and generalizable 3DGS that use an epipolar transformer.
• Insufficient comparative experiments. As the authors mention, this work introduces a novel supervision method, SGC, to enhance the performance of generalizable 3D reconstruction. However, in the NeRF-based backbone experiments, SGC supervision is only added to one method, GNT [1]. I strongly recommend conducting experiments on more generalizable NeRF methods to validate the effectiveness of SGC, such as GNT follow-up methods like EVE-NeRF [2] and GNT-MOVE [3], as well as the transformer-based approach GPNR [4].
• Regarding the concern of "converge too early." In L522, "it should not converge too early when the backbone doesn't have an overall 3D reasoning ability," the solution proposed is to "control the speed of convergence by a small weight." Would it be possible to introduce SGC midway through training? Adding relevant experiments and analysis could be beneficial.

[1] Wang P, Chen X, Chen T, et al. Is Attention All That NeRF Needs?[J]. arXiv preprint arXiv:2207.13298, 2022.

[2] Min Z, Luo Y, Yang W, et al. Entangled View-Epipolar Information Aggregation for Generalizable Neural Radiance Fields[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 4906-4916.

[3] Cong W, Liang H, Wang P, et al. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023: 3193-3204.

[4] Suhail M, Esteves C, Sigal L, et al. Generalizable patch-based neural rendering[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022: 156-174.

1. The improvement over existing work is marginal. The method deals with very similar issue as NeuRay, i.e. how to aggregate the samples in a ray according to their visibility or importance to the novel views. However, when compared with NeuRay, the improvement is very marginal or worse on some metrics. Also, why the comparison on DTU dataset in Table 1 is missing?
2. The ablation of adding the proposed loss on the GNT backbone in Table 1 also demonstrates the very marginal improvement, making it suspectable as evidence of the proposed method's effectiveness. Though in Table 2, in large baseline, the method demonstrates large improvements over pixelSplat, it does not show whether on GNT, similar improvements can be achieved. Especially, compared with NeuBay, can the method demonstrate similar improvements?
3. To consolidate the work, IBRNet+ SGC and NeuRay+SGC are suggested.

1. The technical novelty is not sufficient, as the key technical contribution, which is a change in the optimal condition of depth regularization proposed by Mip-Nerf 360, lacks significant innovations.
2. The improvement of the method is limited. Although this method produces better reconstruction results in certain areas, it does not achieve results comparable to the baseline in other areas. As shown in Fig. 6, the roof of the house is more blurred compared to PixelSplat, suggesting unstable improvement.
3. The experiments in the paper are insufficient. The paper should include more combinations with different baselines to demonstrate the stability of quality improvements better. The paper also conducts extensive analysis of the incorrect correspondence between viewpoints and proposes to supervise on-scene geometry. The paper should present results under more input-output scenarios and additional geometric visualization results to effectively demonstrate the method's effectiveness.