GaussianCut: Interactive segmentation via graph cut for 3D Gaussian Splatting

We introduce GaussianCut, a new method for interactive multiview segmentation of scenes represented as 3D Gaussians. Our approach allows for selecting the objects to be segmented by interacting with a single view. It accepts intuitive user input, such as point clicks, coarse scribbles, or text. Using 3D Gaussian Splatting (3DGS) as the underlying scene representation simplifies the extraction of objects of interest which are considered to be a subset of the scene's Gaussians. Our key idea is to represent the scene as a graph and use the graph-cut algorithm to minimize an energy function to effectively partition the Gaussians into foreground and background. To achieve this, we construct a graph based on scene Gaussians and devise a segmentation-aligned energy function on the graph to combine user inputs with scene properties. To obtain an initial coarse segmentation, we leverage 2D image/video segmentation models and further refine these coarse estimates using our graph construction. Our empirical evaluations show the adaptability of GaussianCut across a diverse set of scenes. GaussianCut achieves competitive performance with state-of-the-art approaches for 3D segmentation without requiring any additional segmentation-aware training.

Paper

References (65)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer MadC8/10 · confidence 3/52024-06-25

Summary

The paper proposes a method for interactive multiview segmentation of scenes represented as a set of 3D gaussians (3D gaussian splatting). The method takes as an input multiple views of a scene, constructs 3D Gaussian splatting representation of the scene, takes interaction from the user (scribbles, points, or text prompts), uses a video segmentation method (prior work) to get object masks from user input in each view, and then constructs a graph with nodes corresponding to the gaussians. A binary energy function on this graph is designed which encourages similar gaussians (both in space and color) to be connected with edges of higher weight. Also for each Gaussian there is a likelihood term representing how likely it is to be foreground/background. Finally this energy function is optimized with a graph cut and gives a separation of the cloud of gaussians into two sets: foreground/background. After separation in 2 sets, foreground/background can be visualized separately using the corresponding set of Gaussians. Experimental results outperform prior work.

Strengths

The proposed method is well motivated. The energy function makes sense.

Weaknesses

Computational cost is not real time. Mapping user input to gaussians could be better described. For someone not familiar with 3D Gaussian splatting prior work, it is hard to understand how it works. An illustration could help. The cluster similarity in unary terms (Eq. 3) seems to have a big impart on the performance, but is kind of ad-hock not so well motivated. The problem might be a result of 'shortcutting' bias in the energy formulation (cuts separating a small component are cheaper). What you are doing here is kind of reminiscent of saying that pixels close to the foreground scribbles should be more likely in the foreground (if I think of an analogy to 2D segmentation).

Questions

Without the mask refinement method (just from user scribbles), by how much the performance goes down?

Rating

8

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are discussed.

Reviewer DeYw4/10 · confidence 5/52024-07-12

Summary

The paper proposes GaussianCut for interactive multi-view scene segmentation using 3D Gaussian Splatting. It first accepts user input (clicks, scribbles or text, similar to SAM) on single images, and them aim to segment the corresponding 3D Gaussians. The method constructs a graph based on scene Gaussian, and then using graph-cut algorithm to minimize the energy function. Segmentation tracking method are employed to provide the initial segmentation masks. The experiments are performed on LLFF, Shiny, SPIn-NeRF and 3D-OVS.

Strengths

1. The paper has a good presentation and clear writing, where figures are clean for the readers to understand the method pipeline. 2. The paper has a good motivation on extracting objects from explicit 3D Gaussians. Graph cut is well known for image segmentation on pixels, and the extension to 3D Gaussians is interesting. 3. The quantitative results in Table 1, 2, 3 and 4 show the reasonable segmentation performance brought by GaussianCut.

Weaknesses

1. Although extracting objects from explicit 3D Gaussians has a good motivation, this has been studied in previous works like LangSplat [a] and Gaussian Grouping [b], but these two very relevant methods are missing quantitative comparisons in the main paper. It is not clear what are the advantages of the proposed Gaussian Cut comparing to [a] and [b]. In Table 10 (should move to the main paper), [a][b] are compared but the performance of Gaussian Grouping is still on par or even better than the proposed GaussianCut. To understand the paper novelty, the paper should clarify these main differences / advantages in the main paper, and also include the detailed running speed comparison. 2. Using video tracking masks to obtain coarse segmentation is not new, which has been explored in [b, c], but not compared. Since [a][b] both lift SAM's masks to 3D, these 2 methods can also perform click/scribble based 3D segmentation 3. The limited performance improvement of the introduced spatial, color and cluster similarities in Table 5, which show the improvement of the proposed n-links and t-links are minor. Also, the extension of Graph cut from 2d pixels to 3D Gaussians seems very straight-forward. 4. More comparisons on benchmarks like Replica (setting proposed by Panoptic Lifting [d]) or Lerf-Mask [b] are desired. [a] Langsplat: 3d language gaussian splatting. CVPR, 2024. [b] Gaussian Grouping: Segment and Edit Anything in 3D Scenes. ECCV, 2024. [c] CoSSegGaussians: Compact and Swift Scene Segmenting 3D Gaussians with Dual Feature Fusion. arXiv, 2024. [d] Panoptic Lifting for 3D Scene Understanding with Neural Fields. CVPR, 2023.

Questions

How will GaussianCut perform when video segmentation/track fail due to large motion? which may lead the coarse segmentation contain large portions of errors. How many Gaussians are considered during the graph construction? The most concerning part for me is the tech novelty of the paper.

Rating

4

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

The limitation is discussed in the paper.

Reviewer 24kC6/10 · confidence 3/52024-07-12

Summary

This paper presents GaussianCut, a new method for foreground/background segmentation of 3D Gaussian scenes. GaussianCut relies on 2D image/video segmentation masks, which are generated for a subset of the training images. GaussianCut propagates these masks to the 3D Gaussians by projecting each Gaussian onto each mask and averaging the corresponding mask values. It then assembles the Gaussians into a graph where each Gaussian is connected to its nearest neighbors, defines several energy terms (for nodes and edges), and uses these energy terms alongside a graph cut to partition the Gaussians. Unlike several baselines, GaussianCut can be applied to pre-trained 3D Gaussian scenes without retraining/fine-tuning.

Strengths

- In Figure 4, the qualitative results of the proposed method look much sharper than those of the baselines. Also, the quality metrics for the proposed method improve over the baselines slightly. - Figures 5 and 7 make it clear that the proposed method is fairly robust to bad 2D input segmentations. - The paper is clear and easy to follow. - The proposed method is simple.

Weaknesses

- One major weakness is that none of the tables include the runtimes of the proposed method and the baselines. This is important because segmentation is arguably much more useful if it runs at interactive speeds. The paper also does not list the runtime of the simpler coarse splatting variant/ablation. I think it is crucial for the paper to list both the training-time overhead and segmentation runtime of each method. - Another major weakness is that the paper includes/omits baselines in the tables very inconsistently. For example, SAGD and ISRF are featured in Figure 4, but not in Table 3. On the other hand, MVSeg and SAGA are featured in the table, but not the figure. I would be much more confident in the results if the baselines were always included (when appropriate). Overall, I would be happy to raise the paper's score if the above two major weaknesses were addressed. My score for contribution is held back by the fact that some of the baselines (e.g., SAGA) are significantly faster to run despite having similar quality. - Another smaller weakness is that the paper does not include visualizations of the energy terms, high-confidence clusters, etc. These would be very helpful for building the reader's intuition for the method. Minor nitpicks: - Line 55/56: The following sentence should perhaps include citations: "Recent works have also explored segmentation with Gaussians as the choice for scene representation." - Line 101: I was confused about what is meant by "decomposing the boundary Gaussians." - Line 144: "Guassians" - Line 169: The graph is defined as $(|\mathcal{G}|, \mathcal{E})$. This should be $(\mathcal{G}, \mathcal{E})$ instead. - Line 170: The definition of the neighborhood is listed as $\mathcal{N} \subseteq |\mathcal{G}| \times |\mathcal{G}|$, but the neighborhood is defined in terms of nodes ($\mathcal{N} \subseteq \mathcal{G}$) and not edges (which would be $\mathcal{N} \subseteq \mathcal{G} \times \mathcal{G}$ anyways) in the next sentence. - Line 264: Units (dB) should be listed for the PSNR differences. - Line 296-297: The time cost for segmentation technically grows linearly, but the constant factor is big enough that this doesn't matter. I would update the sentence here to be more precise. - The single mask ablation (line 298) should probably be included in Table 5.

Questions

- Line 171-172: The authors state that "Gaussians that map to the same object would be closer spatially." This seems reasonable, but isn't always the case. For example, dull specular highlights are often represented via transparent surfaces and "clouds" inside objects. Does the proposed method handle these cases well? - The proposed method has a number of hyperparameters. How sensitive is it to these hyperparameters? - How good is the coarse splatting baseline with well-chosen hyperparameters? I think visualizing a sweep of the threshold in the appendix would convince the reader that the proposed approach works better no matter what threshold is chosen. - How are the frames ordered before they are passed to the video segmentation model? How sensitive is the method to this ordering? - To what extent does the proposed method's performance rely on SAM-Track's quality? Do any baselines rely on different video segmentation models, and if so, how do the metrics change when those are updated to use SAM-Track?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Reviewer BR3k6/10 · confidence 4/52024-07-13

Summary

This paper proposes GaussianCut for interactive 3D segmentation. The GaussianCut takes a trained 3DGS model and the user prompt as inputs. The SAM model first transfers the user prompt into an initial mask. Then the 3DGS model renders multiple view images and an existing video-tracking model is used for segmenting 2D masks across multiple views. With the masks on multiple views, the splatted Gaussians are identified with two likelihood parameters. Then the graph-cut method is applied for the Gaussian points where each Gaussian is a node, and the edge models the foreground and background relations. Results on multiple benchmarks show the effectiveness of the proposed method.

Strengths

- This paper is well-written and easy to follow. - Using Graph Cut to segment the 3D Gaussians makes sense and sounds interesting. - The result on multiple benchmarks shows the proposed method achieves new SOTA performance.

Weaknesses

- The proposed method is kind of straightforward. The framework of GaussianCut is a combination of all existing models. The SAM model is used to obtain masks from user prompts. The SAM-Track model is used to generate Multiview masks. 3DGS model can explicitly model relations between 3D Gaussians and 2D pixels. Graph Cut model is used to separate the Gaussian points. This paper should explain the key contributions of GaussianCut itself. - Details of associating 3D Gaussians with 2D masks should be given. As alpha blending will assign different weights to different Gaussians for a single pixel, For the Gaussians splatted into one pixel, are they assigned different weights just for this pixel? Or a hard (binary) assignment is used? - The time cost of the proposed method is shown in Tables 6 and 7. However, the comparisons with other SOTA methods in terms of time cost should be included to illustrate the speed advantage/disadvantage of the proposed method.

Questions

The novelty of the proposed method should be justified. The proposed method is more like a post-processing. Some important method details and comparisons are missing.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have adequately addressed the limitations.

Authorsrebuttal2024-08-13

We thank the reviewer for the follow-up. > Coarse splatting We elaborate on the coarse splatting in further detail below: Consider an optimized 3DGS model for a scene, $\mathcal{G}$. For $n$ viewpoints, we obtain masks $\mathcal{M}$ := {$M^j$}$_{j=1}^n$ from a video segmentation model corresponding to $n$ viewpoints $\mathcal{I}$ := {$I^j$}$_{j=1}^n$ . $M^j$ indicates the set of foreground pixels in the viewpoint $I^j$. For each Gaussian $g \in \mathcal{G}$, we maintain a weight, $w_g$, that indicates the likelihood of this Gaussian belonging to the foreground. To obtain the likelihood term $w_{g}^j$ pertaining to mask $j$ for Gaussian $g$, we unproject the posed image $I^j$ back to the Gaussians using inverse rendering and utilizing the mask information, $$ w_g^j = \frac{\sum_{\textbf{p} \in M^j } \sigma_g(\textbf{p})T_g^j(\textbf{p}) }{\sum_{\textbf{p} \in {I}^j} \sigma_g(\textbf{p})T_g^j(\textbf{p})} $$ where $\sigma_g(\textbf{p})$ and $T_g^j(\textbf{p})$ denote the opacity and transmittance from pixel $\textbf{p}$ for Gaussian $g$. If $g$ does not contribute to $\textbf{p}$, the transmittance is taken to be $0$. Combining over all the masks, $$ w_g = \frac{\sum_{j}\sum_{\textbf{p} \in M^j } \sigma_g(\textbf{p})T_g^j(\textbf{p})}{\sum_j\sum_{\textbf{p} \in {I}^j} \sigma_g(\textbf{p})T_g^j(\textbf{p}) } $$ As mentioned in the paper, we use the same formulation as proposed by GaussianEditor [a] and kindly refer the reviewer to their paper for further details. For binary weights (hard assignment), we simply keep a count of the number of foreground and background pixels the Gaussian $g$ splats to in $I^j$. $$ w_g = \frac{\sum_{j}\sum_{\textbf{p} \in M^j } \mathbb{I}(T_g^j > 0) }{\sum_j\sum_{\textbf{p} \in {I}^j} \mathbb{I}(T_g^j > 0)} $$ This $w_g$ is used directly in Equation 3 in the paper. We show the IoU on several scenes comparing soft assignments and hard assignments. Since soft assignment has marginally better performance, it is our default implementation. | Scene | Soft assignment | Hard assignment | |--------------------|-----------------|-----------------| | Fern (NVOS) | 83.06 | 82.56 | | Fortress (NVOS) | 97.94 | 98.12 | | Leaves (NVOS) | 95.95 | 95.60 | | Lego (SPIn-NeRF) | 89.18 | 88.95 | | Pinecone (SPIn-NeRF)| 91.89 | 91.99 | > Novelty concerns Regarding the novelty, the graph construction proposed in our method is significantly different from the point cloud literature. Our proposed method is much simpler and both the approaches mentioned by the reviewer involve training networks to obtain the segments. - GrowSP: Unsupervised Semantic Segmentation of 3D Point Clouds: This method does not employ graph cut for segmentation. It learns per-point features, extracts superpoints, progressively grows those superpoints, and performs clustering for segmentation. - SAM-guided Graph Cut for 3D Instance Segmentation: While this method does use graph cut on point clouds, their formulation differs significantly from our approach. They apply graph cut on superpoints which are obtained from another pre-segmentation model. Therefore, when an object to be selected is part of a superpoint, this technique is unable to segment it out. Our approach, on the other hand, provides finer user control as we do not rely on any point cloud pre-segmentation module. Their primary method involves training a graph neural network (GNN) using high-quality pseudo labels whereas our approach is training-free. Their non-GNN baseline also requires masks from multiple views to assign edge weights. Our proposed approach can work even with a single mask. Moreover, since this method stores SAM features, adapting it for 3DGS would require a much higher memory footprint than our approach (please see discussion section). More generally, while some advancements in point cloud segmentation can be tranferrable for 3DGS, our coarse splatting (section 3.3) and terminal links (section 3.4) are heavily tailored for 3DGS. [a] GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting, CVPR 2024

Reviewer BR3k2024-08-13

Thanks for the detailed response. The reviewer appreciates it. Please include those details in the revised version. I will increase my rating accordingly.

Authorsrebuttal2024-08-14

We thank the reviewer for the positive feedback and for expressing the intention to increase the rating. We appreciate the constructive review and will ensure that the discussed changes are included in the revised version of the paper.

Reviewer Qza24/10 · confidence 5/52024-07-13

Summary

The paper presents a method for segmentation in 3D Gaussian Splatting. Given a prompt in 2D, the proposed approach is capable of segmenting objects of interest from the 3D Gaussians. Specifically, the method first performs 2D segmentation in all training views, which are then propagated into 3D using a technique similar to visual culling. This achieves a coarse 3D segmentation in the initial stage. Subsequently, the method employs graph cuts to refine the 3D segmentation. The paper validates the method on multiple datasets using 2D segmentation metrics and demonstrates the potential of the proposed approach.

Strengths

The paper proposes refining 3D segmentation via graph cuts using carefully-designed energy functions. This step is practical and effective. It’s surprising to see that the method, even without 2D mask supervision, can still outperform baselines like LERF.

Weaknesses

The paper has several weaknesses as below: Writing: - The method is not well-motivated in the introduction section. Specifically, in lines 24-39, the problem defined in the paper is hard to understand. The authors might want to rewrite this section to make the motivation clearer. - Part of the implementation is not clear to me. Specifically, I’m uncertain if the method is training-free for segmentation when provided with a pretrained GS model. If so, it would be beneficial to make this clear, as being training-free is a significant strength compared to other methods. Method: - I’m not entirely convinced by the training-free approach in the proposed method. In my view, learning a per-Gaussian feature for segmentation offers more flexibility. Therefore, I suggest the authors justify their method by demonstrating what the proposed method can achieve that learnable features cannot. Experiments: - The important comparison between baselines that use 3DGS such as GaussianEditor and LangSplat. The paper only reports the baselines that use NeRF, which is generally worse than 3DGS in the segmentation task. In other words, the baselines are insufficient to justify the proposed method. - Both qualitative and quantitative results of 3D segmentation are missing in the result section. From my viewpoint, it’s essential to this paper since the proposed method aims to perform 3D segmentation. While 2D segmentation partially shows the performance of the proposed method, it’s not the correct metric to validate the proposed method. The paper should not avoid this evaluation because 3D GT segmentation is missing in the datasets used in the paper. In this case, I would suggest the author trying some other datasets like ScanNet where 3D GT is available.

Questions

Please address the concerns that I described above. During rebuttal, I would expect author justify better the proposed methods according to my suggestions I made above. Before, I currently vote for reject.

Rating

4

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, they did it well.

Reviewer DeYw2024-08-12

I read and appreciate the authors' response to my review. After thoroughly considering the feedback from the other reviewers, I am inclined to uphold my original score as "borderline reject" due to the paper's unclear tech contributions in the submitted paper writing and limited performance improvement. To understand the paper novelty, in the next revision, the paper should highlight / clarify the main differences / advantages to existing works in the main paper, and also include the detailed running speed comparison in the paper as well.

Authorsrebuttal2024-08-12

We thank the reviewer for their suggestion. As per the reviewers' suggestion, we did a detailed run-time comparison against the feature learning methods during rebuttal. We initially did not compare the runtime against feature-based methods as we operate directly on a pre-trained 3DGS model. Since none of the published prior work do a training-free segmentation of 3DGS model, we did not have a direct baseline to compare against. To make our contribution clear, we have highlighted that our approach is training-free in the abstract (line 15), in Figure 1, in the method section (lines 125-126) and the discussion section (lines 319-320). We have differentiated our work from feature-based methods (lines 84-85 and lines 99-100) in the related work section. While we work on improving our work, we were curious to know what other running speeds the reviewer is interested in seeing beyond the ones provided in this discussion period? Regarding the limited performance improvement, we would again like to highlight that for a training-free approach to perform on par or even better than training-based baselines is not only novel but also surprising. As for improvement, we show a considerable improvement on 3D-OVS dataset (+6.88 absolute IoU gain) and a smaller improvement on NVOS (+1.6 absolute IoU) and SPIn-NeRF (+0.5 absolute IoU). The reason for this is that the performance on the two latter benchmarks is already pretty high even for the baselines and thus, the room for improvement is smaller as this benchmark is becoming saturated.

Reviewer BR3k2024-08-12

Thanks for the detailed response. The reviewer still retains the following concerns: 1. Novelty. This paper extends the graph cut-based segmentation method into 3D Gaussians, which haven't been explored before. However, 3D Gaussians are essentially a bunch of 3D points. Using Graph cut for 3D segmentation including point cloud segmentation has been explored [1,2]. The current technological contribution is limited. 2. It's still unclear how the soft/hard weights are combined in the proposed method. Also, what are the experimental results of these two different choices? [1] Zhang, Zihui, et al. "Growsp: Unsupervised semantic segmentation of 3d point clouds." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [2] Guo, Haoyu, et al. "Sam-guided graph cut for 3d instance segmentation." arXiv preprint arXiv:2312.08372 (2023).

Reviewer 24kC2024-08-13

Thank you to the authors for providing detailed runtimes, more baseline results, additional hyperparameter sweeps, and comparisons against the coarse splatting baselines. These additional details adequately address my concerns about the evaluation, and so I have raised my overall score from 5 to 6. I think the paper's direct impact may be limited by its less-than-interactive runtime and the fact that Gaussian segmentation is a very specific niche, but runtime could be improved in follow-up research on training-free Gaussian segmentation, and so I think the paper is worth publishing. I would encourage the other reviewers to consider raising their scores above borderline reject. While the proposed method has the disadvantage of not providing interactive segmentation speed, it takes a fundamentally different approach to segmentation compared to the baselines (training-based vs. training-free), and this is valuable in and of itself. The proposed method has the potential to serve as a stepping stone towards more practical segmentation approaches, and with the additional results the authors have provided, I think the reader will have a good sense of the method's strengths and weaknesses. In other words, although the proposed method is not perfect, I think the paper merits more than a borderline reject.

Authorsrebuttal2024-08-13

Thanks for your kind feedback and for recognizing the value of our work!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC