FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training

The field of novel view synthesis from images has seen rapid advancements with the introduction of Neural Radiance Fields (NeRF) and more recently with 3D Gaussian Splatting. Gaussian Splatting became widely adopted due to its efficiency and ability to render novel views accurately. While Gaussian Splatting performs well when a sufficient amount of training images are available, its unstructured explicit representation tends to overfit in scenarios with sparse input images, resulting in poor rendering performance. To address this, we present a 3D Gaussian-based novel view synthesis method using sparse input images that can accurately render the scene from the viewpoints not covered by the training images. We propose a multi-stage training scheme with matching-based consistency constraints imposed on the novel views without relying on pre-trained depth estimation or diffusion models. This is achieved by using the matches of the available training images to supervise the generation of the novel views sampled between the training frames with color, geometry, and semantic losses. In addition, we introduce a locality preserving regularization for 3D Gaussians which removes rendering artifacts by preserving the local color structure of the scene. Evaluation on synthetic and real-world datasets demonstrates competitive or superior performance of our method in few-shot novel view synthesis compared to existing state-of-the-art methods.

Paper

References (48)

Scroll for more · 36 remaining

Similar papers

Peer review

Reviewer oLUf3/10 · confidence 5/52024-06-25

Summary

This paper presents FewViewGS to regularize sparse view 3DGS from unseen viewpoints without relying on pre-trained depth estimators. The main contribution is to reproject the pixels to an unseen view and calculate the losses at corresponding pixels found by image matching. Experiments on LLFF, DTU, and Blender datasets demonstrate the proposed method can achieve competitive or superior performance compared to existing SOTA methods.

Strengths

- The method can exceed existing baselines without relying on extra depth estimators. - Results are good in visualization.

Weaknesses

- The main contribution of this paper shares the same insight with SPARF [1], which also uses image matching to add supervision on unseen views. The main difference is that SPARF uses depth, while this work uses depth, color and semantic features, for which the novelty in this is limited. And also the multi-stage training is just a common composition of " initialization + regular training + refinement". - The proposed equations have too many manual factors and artificial designs that lack mathematical theory (e.g. Eq (7, 8,13)). These all seem to be engineering attempts but not technical contributions. - Experiment is insufficient and unclear. 1) Although a (Rand. Init.) is mentioned at one item in Table 1, I can not find any description of the used initialization for other settings, which may have a very huge influence on the performance. 2) And serving as a geometry reconstruction method, some geometry visualizations like depth are wished to report. 3) As the reprojection operation and semantic feature extraction may cost a lot of time, I'm wondering if the training time would significantly increase. - Performance improvements are somewhat limited compared to current methods. And although this work can be free from some pre-trained models, it has introduced other dependencies like VGG features and RoMa matching. - The paper has some errors, for example: 1) Eq (6) is unclear on the part (C, Z), which is a 2D->3D back projection I guess. 2) The citation information of SPARF [1] (numbered as [34] in the paper) is wrong. [1] Truong, Prune, et al. "Sparf: Neural radiance fields from sparse and noisy poses." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.

Questions

Please clarify the weaknesses, and make additional explanations of the unclear parts.

Rating

3

Confidence

5

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors have not discussed the limitations. This work seems to heavily rely on some manual hyperparameters, and some pre-trained models like VGG features and RoMa matching, which may limit the applicability to more real-world scenes.

Authorsrebuttal2024-08-10

Reply to Reviewer oLUf

We sincerely appreciate your time and efforts in reviewing our responses. We hope that our responses below adequately address your concerns. ***Q1: Clarification of our novelty*** ***A1:*** In this response, we would like to clarify our novelty and design details. **(1) Novel view consistency constraints:** Our method offers a new solution to address overfitting and ensure multi-view consistency in sparse view synthesis by using correspondence priors to generate labels for novel views. Since accurate depth prediction is crucial for reliable view projection, our method projects matched pairs from training views to the novel view and filters out pairs with significant depth discrepancies in the novel view (see Eq 7). This design enhances label reliability over SPARF's single-view projection. The table below shows the results of SPARF's multi-view correspondence loss (MV-Corr) and depth consistency loss (DCons) on 3DGS, which is much worse than our method (PSNR 19.13, SSIM 0.792, LPIPS 0.186). Furthermore, DCons in SPARF shows limited improvement, see Row i and Row ii. In contrast, our depth supervision for novel views in Table 4iii (PSNR 18.17, SSIM 0.736, and LPIPS 0.198) greatly improves the performance of 3DGS. || Method | PSNR | SSIM |LPIPS | | --- | --- | --- | ---| --- | | i| 3DGS | 15.04 | 0.676 | 0.246 | | ii | + DCons | 15.17 | 0.699 | 0.232 | | iii| + MV-Corr | 16.25 | 0.722 | 0.219 | | iv| + MV-Corr + DCons | 16.37 | 0.729| 0.214 | **(2) Multi-stage training:** Common multi-stage training typically involves an additional teacher network or large dataset. In contrast, our approach tailors the multi-stage training to sparse view synthesis, eliminating the need for extra networks or datasets. It alternates between training and novel views, expanding the view information and solving overfitting. As shown in Table 4ix and 4xi, this multi-stage approach improves the PSNR by 4.4\%. **(3) Locality preserving regularization:** Our locality regularization explicitly constrains the color attributes of 3D Gaussians to preserve local smoothness in 3D space. Table 2 demonstrates reduced artifacts and improved accuracy, and Table 4i and 4ii highlight gains in all metrics. **(4) Loss design:** Our proposed loss functions integrate additional constraints directly, which is intuitive and effective to address issues in sparse view synthesis, without increasing inference time. **(5) $L_{sem}$:** Our work introduces the semantic constraint to integrate local constraints and reduce artifacts. The comparisons between Table 4i and 4v show a 3.2\% increase in PSNR and a 4.5\% reduction in LPIPS. Indeed, as discussed in Lines 272-277, the local constraint needs to balance the benefits of local semantics with the risk of missing details. **(6) Other details:** We apply consistent values for $\theta$ and $\theta_{grad}$ in Eqs (9, 10, 11), to avoid too many human factors. Fine-tuning these parameters and exploring better mapping functions could improve results. ***Q2: Details of SFM initialization*** ***A2:*** The code for SFM process will be released, providing all the details. **(1) SfM initialization**: We follow the SfM process provided in FSGS ('tools/colmap\_llff.py' in FSGS code) to generate camera poses and 3D points, using only training views without any post-processing. The choice of training/test views is consistent with FSGS on LLFF and DNGaussian on DTU. Novel views are randomly interpolated between training views. **(2) Visualizations**: Figure S2 shows 3DGS results generated with SfM initialization, a 3-view setting, and size_threshold=None (as used in FSGS and DNGaussian). We will release the model for further clarification. **(3) COLMAP in DTU**: Following FSGS, we use the dense reconstruction in COLMAP to generate 3D points, with only training views, which successfully yields point clouds on all scenes except 'scan110'. They are used to initialize 3D Gaussians. 'Scan110' is initialized randomly. ***Q3: Comparison with other works based on pre-trained depth/diffusion models.*** ***A3:*** We try to clarify the confusion below. **(1) Training time:** We pre-compute and store VGG features and RoMa-matched pairs during data pre-processing, avoiding extra training time and redundant computations. However, pre-trained depth/diffusion models generate results online, making them more computationally expensive than pre-trained Roma and VGG models. **(2) Quality:** Networks using pre-trained depth/diffusion models generate labels for the novel view often ignore the priors in training views and lack multi-view consistency. In contrast, our method projects labels from training views to novel views, ensuring higher accuracy and multi-view consistency. Using a pre-trained depth network for novel views on 3DGS yielded PSNR 16.74, SSIM 0.734, LPIPS 0.214. Our method with depth supervision alone in Table 4iii (PSNR 18.17, SSIM 0.736, and LPIPS 0.198) outperforms this.

Reviewer oLUf2024-08-11

I'll respond to some important problems, but not mean that other parts are good: 1. For **novel view consistency constraints**: According to the response, the authors agree that their work lacks novelty, since the novelty claimed by the authors, eq (7), is just a simple filter mask supported by a manual hyperparameter, while the other parts in "novel view consistency constraints“ is strongly similar to SPARF. on the part of the effect, the provided reproduction experiment of SPARF on 3DGS is not convincing for me, as there is not any explanation about why the performance gap happens. BTW, according to its paper, SPARF can achieve the performance of PSNR 21.01, SSIM 0.87, LPIPS 0.10 on NeRF backbone in the 3-view DTU setting, which can already fully validate the effect of their strategies. Considering the working principle of this work is extremely close to SPARF without significant innovation, it can be considered as an incremental work to transfer SPARF to 3DGS backbone. 2. For **semantic loss**: According to Table 4 vi and ix, the improvement is very marginal. What's worse, when applying an unsatisfactory type of feature, the performance will drop significantly. On the other hand, Despite that the authors try to prove the effect through "The comparisons between Table 4i and 4v show a 3.2% increase in PSNR and a 4.5% reduction in LPIPS", the analysis is partial. It only compares the situations with raw 3DGS. According to the comparisons between Table 4 vi, vii, viii and ix, its effect significantly shrinks to none after other constraints are applied, showing its redundancy. 3. For the **so-called "SFM initialization"**, it shows that the authors lack the basic knowledge in multi-view geometry and the ability of discernment. Following the description and code process of FSGS, the author declared they use "the SfM process ... to generate camera poses and 3D points, using only training views without any post-processing". However, first, there is a significant mistake in FSGS that it calls the process they used as SfM but actually is an MVS method. Second, this process does not estimate any camera poses. So, the initialization method the authors used is actually MVS, rather than SfM they announced in the paper and rebuttal. This problem shows that the authors do not even know what they are actually doing. Considering this problem is to be revealed after the reviewer's twice asks, I'm worrying about this work's quality. 4. For **introducing pre-trained models**: the authors did not directly reply to my questions. First, are RoMa and VGG pre-trained models? Then, can these pre-trained models escape from the problems that "make training time longer and highly dependent on the quality of the pre-trained models"? In other words, is there not any extra time cost in order to make the proposed method as fast as raw 3DGS, especially when needing to use VGG to generate the semantic embedding for novel view k in eq (11) online? Or can it also work well when using models with poor pre-training quality? If not, how can the authors claim that their method is different from previous works in this part?

Authorsrebuttal2024-08-13

Reply to Reviewer oLUf

We sincerely appreciate your time and efforts in reviewing our responses. We hope that our responses below adequately address your concerns. ***Q1: Clarification of our novelty*** ***A1:*** We strongly disagree with the reviewer's statement about novelty. This assertion is not only inaccurate but also a misrepresentation of our previous responses. Our method aims to address the overfitting and maintain the multi-view consistency, as detailed in Section 3.2. While SPARF also aims to mitigate overfitting, a challenge common to sparse view synthesis, the similarity with our method ends there. Our method utilizes the correspondence priors to guide multi-view projection, filtering out outliers and generating reliable pseudo-labels, which have been verified to be crucial. Subsequently, we compute the appearance/geometry/semantic losses with gradient weighting for the novel view. This not only broadens the available view information but also maintains multi-view consistency. Notably, during loss computation, we again utilize correspondence priors to select the minimal loss among each matched pair, which can reduce the effect of large noises. Table 4ix and 4x prove the effectiveness of the minimal operation. Our method diverges significantly from SPARF in both approach and outcomes, delivering superior performance on 3DGS. The depth consistency loss (DCons) employed in SPARF is fundamentally limited by its reliance on single-view projections, where depth inaccuracies lead to flawed projections. The DCons in SPARF's original paper improves PSNR by 0.20 (from 20.81 to 21.01) on NeRF—a result that aligns with our own reproduction on 3DGS. This proves that the single-view projection on SPARF is insufficient to address the overfitting issue on both NeRF and 3DGS. In contrast, our depth supervision for novel views in Table 4iii improves the PNSR by 3.13. ***Q2: Analysis for semantic loss*** ***A2:*** As explained in Lines 271-277 and Lines 441-443, DINOv2 and CLIP down-sample features early in the encoder with a large stride, e.g. 16. It is widely recognized that features extracted with a large stride often result in a loss of detail, which is the reason why the semantic loss with DINOv2 and CLIP obtains poorer performance. Additionally, Figure 3 further illustrates that using DINOv2 and CLIP as feature extractors generates worse results in some boundary and detailed regions. In contrast, our semantic loss using VGG can reduce artifacts and maintain details. ***Q3: Details of SFM initialization*** ***A3:*** We have already explained in our previous response that we initialize the 3D Gaussians with dense reconstruction in COLMAP. This process involves two parts, SfM followed by MVS, and does estimate camera extrinsics during SfM, therefore, the statement that "this process does not estimate any camera poses" is simply incorrect. Training and test views are both served as input for SfM to obtain their poses, then only the training views are used during MVS to get the fused point clouds. The technically sound initialization method name for the shorthand "SfM initialization" would be "Initialization based on the poses estimated by SfM for all views, and the fused point clouds from MVS on training views only", which will need to be shortened when referring to. Thus the initialization method we used is actually SfM+MVS, rather than just MVS as the reviewer suggested. The integrity of our initialization is by no means affected by the name, we ensure the dense point clouds are only from the training views, making it fair and comparable with the baselines. ***Q4: Comparison of training time and results*** ***A4:*** Our method, which incorporates RoMa and VGG, achieves significantly reduced training time while delivering superior performance in sparse view synthesis compared to networks relying on pre-trained depth/diffusion models, such as FSGS and SparseGS. As noted in our previous responses, only the VGG model introduces additional training time. However, as detailed in Table S7 of the uploaded file, our method requires just 5.82 minutes of training—substantially less than the 51.78 minutes needed by SparseGS (which uses a pre-trained diffusion model) and the 10.90 minutes required by FSGS (which relies on a pre-trained depth network). Furthermore, our approach not only enhances efficiency but also surpasses SparseGS and FSGS in performance, as demonstrated in Table 1. While incorporating VGG does involve some additional training time compared to raw 3DGS, the efficiency and speed trade-off is outweighed by the significant improvements in performance. Our method offers considerable gains over both raw 3DGS and methods reliant on pre-trained depth or diffusion models. For example, it improves PSNR by 4.09 dB on DTU and 3.32 dB on Blender compared to 3DGS. In summary, the integration of RoMa and VGG enhances the overall speed and effectiveness of our method, making it a notable advancement over previous approaches.

Reviewer oLUf2024-08-14

The authors alway passby the most important problems but only quibble by mentioning some irrelevant and trivial things, attempting to divert the attention. They can't face the exposed problems but only deny them. Such a response is meaningless. As above, this paper has many serious problems. The quality can not match NeurIPS. I'll lower my rating. BTW, for the initialization part, the author said that "We follow the SfM process provided in FSGS ('tools/colmap_llff.py' in FSGS code) to generate camera poses and 3D points". Thus, do the authors still claim that "This process involves two parts, SfM followed by MVS, and does estimate camera extrinsics during SfM, therefore, the statement that "this process does not estimate any camera poses" is simply incorrect"? See what 'tools/colmap_llff.py' exactly includes. Where is SfM in this file?

Authorsrebuttal2024-08-14

Reply to Reviewer oLUf

We sincerely appreciate your time and efforts in reviewing our responses. We hope that our responses below adequately address your concerns. We believe that the focus on specific terminology, rather than the actual process and results, is not conducive to the evaluation of the work's quality. We have been transparent in our methodology and provided a thorough explanation in our previous responses (we initialize the 3D Gaussians with dense reconstruction in COLMAP.). This step is performed within the COLMAP framework, specifically using the convert.py script in our provided FSGS code to extract the camera poses. Our use of SfM followed by MVS is technically accurate and appropriately implemented in the context of our research.

Reviewer ZPEG6/10 · confidence 4/52024-07-08

Summary

This paper introduces a novel few-shot Gaussian Splatting method for synthesizing novel views. Unlike conventional approaches that rely on pre-trained monocular depth estimation or diffusion methods, the proposed method leverages the matches of available training views to generate novel sample views between the training frames. It employs color, depth, and image feature losses. Additionally, a novel regularization loss is introduced to preserve the local structure of the object. Experimental results demonstrate that the proposed method achieves significant performance improvements in both real and synthetic datasets.

Strengths

The strengths of the paper are as follows: 1. The proposed few-shot Gaussian splatting method for novel view synthesis, which does not rely on pre-trained depth estimation or diffusion models while achieving state-of-the-art performance. It is important to note that pre-trained depth estimation and diffusion models often require large parameters, which can lead to longer training times. 2. A multi-stage training scheme consisting of pre-training, intermediate, and tuning stages. This scheme optimizes the scene representation gradually. The pre-training process aims to obtain a basic representation of the scene using known views. The intermediate process transfers knowledge from known views to novel views while preserving consistency in overlapping regions between known training views. Finally, the tuning process removes artifacts that occur in few-shot scenarios. 3. The consistency loss function, which maintains the similarity of color, depth, and image features between pixels in novel views projected from known training views. This ensures that novel views have similar semantic, color, and depth information. It is worth noting that the loss functions are adaptively weighted to minimize the impact of errors in texture-rich regions. 4. The locality loss function, which maintains color similarity between the Gaussian and its neighboring regions. Accurate rendering results are more likely to occur when color values are smooth between neighborhoods. 5. The paper provides an exhaustive ablation study for each network design decision, leading to a convincing algorithm design.

Weaknesses

The weaknesses of the paper are as follows: 1. The authors claim that relying on depth estimation or diffusion priors requires longer training times, but there is no comparison or ablation study provided to justify this claim. It would be beneficial to perform a comparison between state-of-the-art methods and the proposed method in terms of training and inference time. 2. The paper does not provide a justification for why the proposed method, which relies on matches between training views, performs better than methods that rely on pre-trained depth estimation or diffusion methods. A more in-depth analysis is needed to demonstrate how the proposed method outperforms these approaches.

Questions

Based on the weaknesses, the following questions arise: 1. How do the training and inference times of the proposed method compare to those of state-of-the-art methods? 2. How well does the performance of the proposed method generalize to unseen data? Given that state-of-the-art methods typically employ large pre-trained models for data reconstruction, the proposed method, which relies solely on matches between known views, may encounter difficulties when dealing with unseen regions.

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The proposed method may face challenges when dealing with texture-rich regions that are not visible from the input views.

Reviewer 8CAh5/10 · confidence 5/52024-07-13

Summary

This paper proposes a new method for sparse-view novel view synthesis. It proposes a multi-stage training scheme including pre-training, intermediate, and tuning stages. It introduces pre-trained dense matching models to find pixel correspondences between different-view images and encourage consistency. A Locality Preserving Regularization is proposed to encourage local smoothness.

Strengths

1. The proposed method achieves SOTA performances on the 3-view LLFF dataset, 3-view DTU dataset and 8-view Blender dataset. 2. The proposed method suits both random points and mvs points as shown in Table 1. 3. The paper is well-organized.

Weaknesses

1. The comparisons are not enough. It lacks comparisons on more input views, such as the 6-view and 9-view settings used in FreeNeRF, which is also important to evaluate the proposed method in sparse-view settings. 2. The ablation studies are not enough. There are lots of hyperparameters listed in the implementation details, however, it is unclear how these hyperparameters are selected and how they impact the performance, which is important to evaluate the robustness of the proposed method. 3. The proposed method introduces ore-trained dense matching networks, thus I think it is still similar to those introduced pre-trained depth estimation networks.

Questions

1. I wonder is the proposed method still work with more input views such as 6 views and 9 views.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have discussed the limitations.

Reviewer Eqzk6/10 · confidence 4/52024-07-13

Summary

This paper tackles the problem of few view (or sparse view) 3DGS multistage training with correspondence-driven losses that enforce projected colors, depths, and semantic features (extracted by a pre-trained VGG) are consistent. Contributions are straightforward and geometrically inspired. In addition, the authors propose a locality preservation loss, which enforces color smoothness in neighbouring gaussians.

Strengths

1. Simple yet effective geometrically inspired regularization for few view 3DGS. 2. Strong qualitative and quantitative results. 3. Extensive ablation studies.

Weaknesses

1. Intuitions behind a 3-stage training strategy are not well established. If intermediate training already includes L_pre_training, why is further fine-tuning with L_pre_training only needed?

Questions

See weaknesses

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Ok

Authorsrebuttal2024-08-10

Reply to Reviewer Eqzk

We sincerely appreciate your precious time and efforts in reviewing our paper and responses. We hope that the response provided below adequately addresses your concerns. ***Q1: Is it possible to show those "unmatched pixels during the intermediate stage" to support your claims?*** ***A1:*** We appreciate the review's suggestion on visualizations to better support our claim. We provided visualizations for the matched results in Figure 7. The white regions in Figure 7c and the areas without red lines in Figure 7d represent unmatched pixels, which lack effective supervision during the intermediate stage. We will include detailed descriptions in our revised version. ***Q2: I am also wondering why you did not include an experiment with 1st: 2000 2nd: 7500 Third: 0, for a more direct comparison.*** ***A2:*** We thank the reviewer for the suggestion on the experiment. The setting with '1st: 2000 2nd: 7500 Third: 0' achieves a performance of PSNR 18.37, SSIM 0.782, and LPIPS 0.198. We will integrate these results into our revised version.

Reviewer oLUf2024-08-08

Thanks for the reply from the authors. After reading the rebuttal, this work still has the following problems. 1. Lack of Novelty. Summarized by the authors in the paper, this work has three main contributions: an effective multi-stage training scheme, novel view consistency constraints, and a color smooth regularization loss. First, as in the review, the multi-stage training is just a common composition of " initialization + regular training + refinement", with no independent novelty on this point. And the color smooth regularization loss is just a trick to add local smoothness to reduce some artifacts, which only improved PSNR but not SSIM and LPIPS, demonstrating its limited effect on the overall quality. If these are declared to be the most important novelties in the paper, I can only believe there are no other valuable things to show. And for the novel view consistency constraints, it is still the incremental technique of SPARF even according to the rebuttal. The contribution lies only in some loss designs with no interesting insight, moreover, has too many manual hyperparameters in eq(7,9,10,11) and replaceable mapping functions like $exp()$ in eq(8, 13). Besides, L_sem is less effect according to Table 4 (vi, ix, xiv) and Figure 4. Especially, some details became worse after adding L_sem in Figure 4. It is also sensitive to the type of feature. I do not think these novelties can match the bar at NeurIPS. 2) Details of "SFM initialization" is still not clear enough. Please explain how many views are involved in the SfM process to get the camera poses and point cloud, how to align the excluded test views into the camera coordinate system, and whether is there any post-processing of the point cloud. As far as I know, original 3DGS can not recover such precise scenes shown in Figure S2 when just use 3-view SfM point cloud as initialization, and COLMAP will fail in some DTU scenes when using only three training views. I have doubts about these. 3) The authors claimed introducing depth or diffusion models will "make training time longer and highly dependent on the quality of the pre-trained models" (lines 40-41), which is the stem of this work. However, why introducing pre-trained RoMa and VGG can escape from these weaknesses? The logic is weird. I'll temporarily keep my rating.

Reviewer Eqzk2024-08-09

Thanks for your reply

Thanks for supplementing your results. It is now clear the 3rd stage provides additional performance improvements. Is is possible to show those "unmatched pixels during the intermediate stage" to support your claims? I am also wondering why you did not include an experiment with 1st: 2000 2nd: 7500 Third: 0, for a more direct comparison.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC