XMask3D: Cross-modal Mask Reasoning for Open Vocabulary 3D Semantic Segmentation

Existing methodologies in open vocabulary 3D semantic segmentation primarily concentrate on establishing a unified feature space encompassing 3D, 2D, and textual modalities. Nevertheless, traditional techniques such as global feature alignment or vision-language model distillation tend to impose only approximate correspondence, struggling notably with delineating fine-grained segmentation boundaries. To address this gap, we propose a more meticulous mask-level alignment between 3D features and the 2D-text embedding space through a cross-modal mask reasoning framework, XMask3D. In our approach, we developed a mask generator based on the denoising UNet from a pre-trained diffusion model, leveraging its capability for precise textual control over dense pixel representations and enhancing the open-world adaptability of the generated masks. We further integrate 3D global features as implicit conditions into the pre-trained 2D denoising UNet, enabling the generation of segmentation masks with additional 3D geometry awareness. Subsequently, the generated 2D masks are employed to align mask-level 3D representations with the vision-language feature space, thereby augmenting the open vocabulary capability of 3D geometry embeddings. Finally, we fuse complementary 2D and 3D mask features, resulting in competitive performance across multiple benchmarks for 3D open vocabulary semantic segmentation. Code is available at https://github.com/wangzy22/XMask3D.

Paper

References (52)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer tdkT5/10 · confidence 3/52024-06-25

Summary

This paper introduces XMask3D, a framework developed for open vocabulary 3D semantic segmentation. They propose the integration of the denoising UNet, derived from a pre-trained diffusion model, to generate geometry-aware segmentation masks conditioned on learnable implicit 3D embeddings. These binary 2D masks are used to filter mask-level embeddings of 3D representations and apply mask regularization, thereby improving the open vocabulary capacity of 3D features.

Strengths

1. The motivation is clear. 2. The proposed method is intuitive, and the experiments have validated their contributions.

Weaknesses

1. The organization should be improved. Section 3.1 provides an overview, while section 3.2 includes design insights and preliminary findings. The flow of these writings has puzzled me, making it difficult to grasp your key contribution.

Questions

1. Please provide further clarification on how mask-level alignment between 3D features and 2D embedding space can address the limitations of traditional techniques, such as global feature alignment or vision-language model distillation. Additionally, if texts (Category Labels) are concatenated with fused features, will it still create a unified feature space that encompasses 3D, 2D, and textual modalities? 2. Could you please provide further clarification on the main contributions of your research compared to PLA and OpenScene? Although the 3D caption process shares similarities with PLA, the overall pipeline resembles OpenScene, with the exception of the diffusion model and mask generator, which differ from the Multi-view Feature Fusion in OpenScene. 3. Does the Implicit 3D Captioner effectively work with your 3D features? From my understanding, the most reliable 3D captioner currently available is Cap3D, which generates captions for 3D objects by rendering multi-view images and utilizing BLIP2 and LLM for assistance. In the context of indoor-scenes, can we consider the Implicit 3D Captioner to be equally robust? It would be beneficial to present additional evidence to support this claim. 4. Can your text-to-image diffusion model effectively generalize to your datasets? If not, please provide examples of failure cases. Additionally, is the diffusion model fine-tuned during the training process or is it frozen? If not, please present additional results to demonstrate the robustness of your diffusion model in generating high-quality images within your datasets. 5. What is view-level contrastive loss? Why this loss is calculated between the view global feature and text embedding of the view image caption but have three coefficients? 6. It is recommended to show your 2D Mask and 3D Mask in Figure 3 to provide more visual evidence. 7. The authors should provide results on ScanNet++ (CVPR’23), which is a up-to-date dataset compared with ScanNet. 8. Since diffusion models are utilized, it is recommended to compared the model parameters and FLOPs compared with PLA and OpenScene.

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

1. The authors have addressed the limitations and potential negative societal impact of their work.

Reviewer AnYT5/10 · confidence 4/52024-07-12

Summary

The paper proposes a precise and consistent mask-level alignment between 3D features and the 2D-text embedding space through a method called cross-modal mask reasoning. The proposed XMask3D model includes a 3D branch for capturing geometric features, a 2D branch for generating vision-language aligned masks, and a fusion block to combine 3D with 2D. Using a pre-trained text-to-image diffusion model as the 2D mask generator, the model leverages three techniques: 3D-to-2D mask generation, 2D-to-3D mask regularization, and 3D-2D mask feature fusion.

Strengths

1- The idea is novel, the author propose to merge 2D which provides high OV capabilities, with 3D features shich endoces 3D geometry. 2- The method performs remarkably better than the reported models, namely OpenScene. The experiments are also well structure

Weaknesses

1- The authors don't compare with state-of-the-art 3D semantic segmentation OV3D[1] 2- The authors highlighed fututre work in the limitation, it would be good if you can expand it with some limitation on the technical side or some failure cases. [1] Jiang, Li, Shaoshuai Shi, and Bernt Schiele. "Open-Vocabulary 3D Semantic Segmentation with Foundation Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.

Questions

Please compare to OV3D mentioned in the weaknesses

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Needs to be expanded

Reviewer Xo8p7/10 · confidence 4/52024-07-12

Summary

The paper addresses the limitations of current open vocabulary 3D semantic segmentation methods, which primarily focus on creating a unified feature space for 3D, 2D, and textual modalities but struggle with fine-grained segmentation boundaries. To overcome these limitations, the authors propose XMask3D, a cross-modal mask reasoning framework that achieves more precise mask-level alignment between 3D features and the 2D-text embedding space.

Strengths

1. The part "incorporating a 2D mask generator to create geometry-aware open masks and apply fine-grained mask-level regularization on 3D features" seems reasonable and novel. 2. The paper is well-structured and easy to follow. 3. Analysis is thorough and insightful.

Weaknesses

1. The paper evaluates the proposed method on a limited set of benchmarks (ScanNet20, ScanNet200, S3DIS), all of which are indoor scene datasets. Authors could discuss how the method might perform on outdoor datasets. Additionally, the authors could provide a qualitative analysis of the model's potential limitations when applied to different environments. 2. The reliance on the denoising UNet from a pre-trained diffusion model could be seen as a potential weakness or limitation, especially given the computational resources required for training and inference.

Questions

1. The paper could benefit from a more detailed error analysis to understand the failure modes of XMask3D, especially in novel category segmentation.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have addressed limitations.

Reviewer 4VWC6/10 · confidence 4/52024-07-14

Summary

This paper addresses the challenge of open-vocabulary 3D semantic segmentation by utilizing 3D geometric features, 2D semantic embeddings, and text modality. The proposed approach adapts the ODISE method to the 3D domain, aiming to distill open-vocabulary semantic segmentation knowledge from a pre-trained text-to-image denoising diffusion model to a 3D segmentation model. Initially, an input point cloud is fed into a 3D encoder-decoder segmentation network, producing point-wise geometric features. Simultaneously, a pre-trained visual-language diffusion model generates 2D masks and embeddings from posed images of the same scene, conditioned on the 3D global feature of the 3D branch’s encoder. Unlike the ODISE method, an implicit $3D$ captioner is introduced to produce geometry-aware 2D masks while also distilling information from the 2D branch network to the 3D encoder. To further regularize the 3D network, a distillation loss ($\mathcal{L}_{mask}$) is applied to the 3D mask embeddings, derived from the per-point features and the 2D masks back-projected to the point cloud as 3D binary masks. By obtaining ground truth mask features from a pre-trained CLIP model, the 3D masked embeddings are aligned with the image-text joint embedding space, through a cosine similarity loss. This alignment leads to more coherent segmentation results and enhances the model's open-vocabulary capabilities. Finally, the per-point features are combined with the pseudo mask 2D features (formed by the back-projected 3D mask and 2D mask embeddings), resulting in a fused per-point representation that incorporates the geometric information from the 3D segmentation network and the semantic open-vocabulary capabilities of the 2D branch. The approach is evaluated on three semantic segmentation benchmarks (ScanNet, ScanNet200, and S3DIS) and demonstrates superior performance compared to competing methods.

Strengths

The XMask3D effectively aligns 3D geometric features with 2D and textual modailities through knowledge distillation from visual-text joint embedding spaces inherent in the pre-trained 2D denoising UNet and the CLIP model. As evident by the ablation, the implicit 3D captioner is a crucial step in the overall pipeline, and it outperforms vanilla text conditioning or the implicit 2D captioner of ODISE, in both base and novel semantic categories. Moreover, the 2D-to-3D mask regularization is also essential, since it significantly improves the accuracy of the proposed method esp. in novel categories. This justifies the need for this additional distillation step from the CLIP joint space, to further enhance the open-vocabulary capabilities of the XMask3D method. Finally, the discussion on modality fusion, both in the main paper and supplementary, is highly appreciated. By dissecting the method and providing qualitative and quantitative results for each step, the authors make it easier for readers to understand and gain intuition about the presented approach.

Weaknesses

While the method exhibits superior performance w.r.t. competing methods, it seems that the output fused embeddings yields to geometric inconsistent features for semantic classes that cover large areas of the point cloud such as wall, ceiling and floor. This is evident in both partitioning settings when the class is either base or novel (Table 5 (a) and (b) in supp.).

Questions

Following the weaknesses section, do the authors have any additional insights into why this phenomenon occurs with the fused embeddings? Could the 2D regularization terms ($\mathcal{L}_{seg}^{2D}$, $\mathcal{L}\_{view}^{2D}$) be introducing too much bias towards the 2D visual modality, thereby causing geometric discontinuities in the output fused features?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the authors have discussed the method's limitations in detail in Section 4.4.

Reviewer Xo8p2024-08-08

I've thoroughly reviewed the authors' responses and appreciate their thoughtful engagement. Most of my concerns have been addressed. I will stay in touch for further discussion as we approach the final rating.

Authorsrebuttal2024-08-12

Response to Reviewer Xo8p

We greatly appreciate your response and valuable suggestions, which improved the quality and comprehensiveness of our paper! We will update our revised paper according to our discussions.

Area Chair Eak82024-08-10

Discussions and finalize rating

Dear Reviewers, Thank you for your efforts. Please review the rebuttals, engage in the discussions, and provide your final ratings. Thank you again for your valuable contributions. AC

Authorsrebuttal2024-08-12

Response to Reviewer tdkT

Thanks for upgrading your score and providing valuable feedback. We will update our revised paper according to our discussions and release our code and checkpoints after the conference decision. Thank you again for your insightful and constructive suggestions that improve paper quality!

Reviewer AnYT2024-08-12

Thanks a lot for clarifying these points. The authors addressed my comments. Thus, I am happy to keep my score, and it would be great to report the results of OV3D and XMask3D together in the revised version.

Authorsrebuttal2024-08-13

Response to Reviewer AnYT

We deeply appreciate the time and effort you dedicated to the careful review and insightful feedback on our paper! We will update our revised paper and add OV3D comparison according to our discussions.

Reviewer 4VWC2024-08-14

Thank you for the clarifications and for addressing my concerns. I will maintain my positive rating

Authorsrebuttal2024-08-14

Response to Reviewer 4VWC

We greatly appreciate your response and valuable suggestions, which improved the quality and comprehensiveness of our paper! We will update our revised paper according to our discussions.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC