Bridging the Domain Gap: Self-Supervised 3D Scene Understanding with Foundation Models

Foundation models have achieved remarkable results in 2D and language tasks like image segmentation, object detection, and visual-language understanding. However, their potential to enrich 3D scene representation learning is largely untapped due to the existence of the domain gap. In this work, we propose an innovative methodology called Bridge3D to address this gap by pre-training 3D models using features, semantic masks, and captions sourced from foundation models. Specifically, our method employs semantic masks from foundation models to guide the masking and reconstruction process for the masked autoencoder, enabling more focused attention on foreground representations. Moreover, we bridge the 3D-text gap at the scene level using image captioning foundation models, thereby facilitating scene-level knowledge distillation. We further extend this bridging effort by introducing an innovative object-level knowledge distillation method that harnesses highly accurate object-level masks and semantic text data from foundation models. Our methodology significantly surpasses the performance of existing state-of-the-art methods in 3D object detection and semantic segmentation tasks. For instance, on the ScanNet dataset, Bridge3D improves the baseline by a notable margin of 6.3%. Code will be available at: https://github.com/Zhimin-C/Bridge3D

Paper

Similar papers

Peer review

Reviewer MtnQ5/10 · confidence 4/52023-07-06

Summary

The paper proposes a pre-training strategy to distill the knowledge for large image foundation models. The method proposed is based on a mask auto-encoder transformer model and the setup contains several modules or losses. First, the masking strategy is guided by segmentation labels obtained using foundation models. Moreover, two additional losses at the scene and object level are incorporated to distill information from image foundation models. Lastly, a loss is used to align features from the point encoder to a text encoder.

Strengths

The paper proposes a combination of several foundation models with different modalities, such as segmentation masks, unsupervised pre-trained features, and text embeddings. Moreover, the paper is easy to follow.

Weaknesses

- The novelty of the methods is questionable since other methods have used text foundation models (Language-Grounded Indoor 3D Semantic Segmentation in the Wild), and image pre-trained models (OpenScene: 3D Scene Understanding with Open Vocabularies), or both (ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D Understanding). The paper proposes to combine similar ideas, but they are not new and not enough comparisons to those other pre-training strategies are provided. - To use an informed mask guidance has been proposed before and the paper misses to mention all related research in this area. Moreover, the proposed masking method is not particular to 3D data and these other strategies for images could be used instead. These papers should be cited and compared to these approaches. SemMAE: Semantic-Guided Masking for Learning Masked Autoencoders What to Hide from Your Students: Attention-Guided Masked Image Modeling - The evaluation in general is limited and the only conclusion the reader can draw from these results is that Bridge3D is a better pre-training strategy than vanilla Point-MAE (which is not surprising), and PiMAE. The other comparisons provided in the table are comparing apples and oranges. Most of the methods in Table 1 use VoteNet as an object detection framework, a much older framework than the ones used with Bridge3D. Is the improvement coming from the pre-training or from the newer detection frameworks? Moreover, some of the numbers in the table are not correct. For example, DepthContrast provides results for two object detection frameworks and architectures, VoteNet and H3DNet. However, only the results for VoteNet are presented which are much lower than H3DNet (64.0/42.9/61.6/35.5 vs 69.0/50.0/63.5/43.4). Moreover, these methods use different architectures which makes the pre-training strategies not comparable. In order to claim superiority over all these pre-training strategies the same model and framework for object detection should be used. Since, the proposed method is based on MAE and therefore limits the architecture choice, at least the performance of the proposed method with a VoteNet framework (and other methods with the newer object detection frameworks) should be reported. Regarding semantic segmentation, the evaluation has also limitations. First, only results for S3DIS are provided. This data set is small and relatively large models overfit very easily, most of the recent improvements in the data set are obtained by the type of data augmentations used. I would suggest the authors to report results on a large data set such as ScanNet, which, although not perfect, is more representative of this task. Moreover, there are other works missing in Table 2 that appear in Table 1, for example [27]. They report 72.2 mIoU which is higher than the one obtained with the proposed method, Bridge3D.

Questions

I do not have any questions for the authors. I believe the paper needs to significantly increase the evaluation to show which is the real contribution wrt previous work and to support the claims of the paper.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

1 poor

Limitations

The paper does not analyze properly the limitations of the method. For example, authors could discuss: some of the foundation models used are trained with annotated segmentation masks. Does this pose a problem for the test sets used on the paper? how similar are the images used for training these foundation models to the images used during pre-training and downstream tasks?

Reviewer hzsQ6/10 · confidence 4/52023-07-06

Summary

This paper aims to improve the pre-training of 3D point cloud models by leveraging various foundational models, including vision models such as SAM and Grounding Dino, language models like ChatGPT, and vision-language models. To achieve this objective, the paper proposes several techniques: 1. Generating instance segmentation masks for the corresponding 2D images of 3D point clouds. This is accomplished by first generating image captions using caption models and language models like Tag2text and ChatGPT. Then, vision models such as SAM and Grounding Dino are employed to generate masks based on these captions. 2. Masked point cloud reconstruction with the masks obtained in the previous step. 3. Scene-level distillation using the features of an image encoder (2D-3D Disti) and a language encoder (Text-3D Distill). 4. Object-level distillation using the features of an image encoder (2D-3D Disti) and a language encoder (Text-3D Distill). The proposed pre-training method is evaluated on several downstream tasks, including ScanNetV2 and SUN-RGBD object detection, as well as S3DIS semantic segmentation.

Strengths

1. The idea of utilizing foundational models for 3D point cloud modeling is both interesting and promising. 2. The paper is well-written and easy to follow. 3. The experimental results consistently demonstrate improvements in several downstream tasks when compared to the baseline.

Weaknesses

1. My main concern is that the experimental results are not sufficiently convincing. Although the results demonstrate consistent improvements compared to the baselines, the performance of the baselines used in this paper appears to be relatively low (e.g., the results of 3DETR and GroupFree3D). Additionally, the final system's performance lags behind the current state-of-the-art methods. 2. The overall pipeline is complex, making it difficult to evaluate the effectiveness of each component, such as the model selection for each component. 3. typos: (1) On line 159, the reference 'MAE [63]' is incorrect. (2) On line 233, it should be 'text branch' instead of 'image branch' when mentioning the utilization of CLIP ViT-B [48] to extract text features. (3) The teaser of Figure 2 is somewhat confusing.

Questions

See Weaknesses

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

See Weaknesses

Reviewer E1Dc6/10 · confidence 4/52023-07-06

Summary

a novel masking and patch-dropping strategy based on foundation models to refine the focus of the network on foreground representation learning for 3D masked autoencoders. a novel scene-level and object-level multi-modality knowledge distillation method that pre-trains a 3D network via features, semantic masks, and captions obtained from foundation models

Strengths

Personally I really like the self-supervised framework. It's natural to use MAE to reconstruct and build 3D representations, and obtaining 3D features from 2D images and text is also reasonable. The segmentation results seem good.

Weaknesses

* Lack of qualitative examples. The biggest problem to me is that no qualitative examples are presented in the paper (and visualization is really important for 3D reconstruction and segmentation!) * The experiments focus on segmentation. However, there's more to do with foundation models - like showing zero-shot results, captioning (generation or retrieval) etc. * The paper doesn't mention several recent papers on 3D representation / feature / concept learning from 2D features (like CLIP). Just to name a few: ConceptFusion: Open-set Multimodal 3D Mapping Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models 3D Concept Learning and Reasoning from Multi-View Images This paper is also really really similar: ULIP: Learning a Unified Representation of Language, Images, and Point Clouds for 3D Understanding Also, I think the self-supervision and feature learning framework shares merits with this paper: 3D Concept Grounding on Neural Fields. Some of these methods could serve as baselines for segmentation. If not, you *should at least cite a few papers in the domain*.

Questions

Misc: The *Mask*s mentioned in 3D representation and in 2D segmentation are two different things and may seem confusing. It's better if you could replace one of them by other words (like occupancy for 3D)

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

see weaknesses

Reviewer XLcq6/10 · confidence 4/52023-07-06

Summary

This paper proposes a combination of several techniques to pre-train 3D models by leveraging 2D foundation models. Specifically, the paper proposes: a) an MAE pre-training guided by foreground masks generated by foundation models to mask the foreground masks more compared to the background, to encourage the models to focus more on the foreground. b) Scene-level feature distillation between 3D-image and 3D-text with the help of image captioning foundation model c) Object level feature distillation with the help of mask generated via foundation models. The results show better performance compared to prior methods of self-supervising the 3D models and the ablations confirm the benefit of each proposed technique.

Strengths

The paper is overall well-written and the results are impressive compared to the baselines. The ablations are clear and well-motivated. Investigating how far we can get by leveraging 2D foundation model is an important research direction, and this paper does a good job in exploring several ways to go about it.

Weaknesses

I do not have any substantial complaint with this paper, however, I think the main limitation is in the complexity of the proposed technique: the approach has several moving components relying on several off-the-shelf models and it seems like a combination of a bunch of techniques to utilize the 2D foundation models; requiring several offline computation of intermediate results.

Questions

N/A, I do any have any specific questions.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No, I think it would be useful to specify limitations in the main paper (for eg: complexity of the approach)

Reviewer V9nH5/10 · confidence 4/52023-07-07

Summary

This paper proposes a methodology named Bridge3D to pre-train 3D models using features, semantic masks, and captions sourced from foundation models. Experiments have been conducted on different 3D understanding frameworks.

Strengths

The overall architecture seems to be reasonable; using the modern foundation models and multimodal supervision is promising to help 3D understanding, and some of the results show a good improvement gap compared to training from scratch.

Weaknesses

This framework seems to be overall complicated, which might involve quite some tuning effort and impact its scalability.

Questions

The reviewer is curious whether the authors have tried to simplify this framework. Another question is that, for the 3D object detection results, table 1, it seems only the basic versions of the models have been tried. For example, the GroupFree3D, the best performance on ScanNetV2 reported in this paper is 69.7 and 52.4, however, in the GroupFree3D original paper, the best performance they reported is 69.1 and 52.8, which is basically on par with the reported results in this paper, the 3DETR results have the same issue. Have the authors tried to apply your method to their most powerful models? What is the highest performance you may get in that case, will that reach new SOTA results? So far the improvement compared to the vanilla baseline looks fine but in order to demonstrate the proposed method to be more useful, the absolute performance or the highest performance by applying the proposed method needs to be explored.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Refer to the weakness part.

Authorsrebuttal2023-08-15

**Extra response to the concern on complexity.** As demonstrated in Table 1, the results indicate that each modality contributes to better results. When we combine all modalities, our proposed method achieves the best result. Furthermore, our method demonstrates robustness to variations in the choice of modalities. This flexibility in modal choice means that the complexity of the framework can be scaled down without significant degradation in performance. In essence, users can tailor the framework to their specific needs, allowing for a more adaptable application of our method without sacrificing quality. **Table 1: The effectiveness of each modality. Ablation study on the effectiveness of each modality on 3D object detection and semantic segmentation tasks**. | Image Modality | Text Modality | ScanNetV2 $AP_{25}$ | ScanNetV2 $AP_{50}$ | S3DIS $mIoU$ | S3DIS $mAcc$ | |----------------|---------------|---------------------|---------------------|--------------|--------------| | | | 62.1 | 37.9 | 61.1 | 67.2 | | ✔️ | | 65.7 | 44.6 | 69.0 | 75.4 | | | ✔️ | 64.8 | 43.2 | 68.4 | 75.2 | | ✔️ | ✔️ | **66.3** | **45.5** | **71.1** | **77.5** |

Authorsrebuttal2023-08-15

**Extra response to the concern on complexity.** As demonstrated in Table 1, the results indicate that each modality contributes to better results. When we combine all modalities, our proposed method achieves the best result. Furthermore, our method demonstrates robustness to variations in the choice of modalities. This flexibility in modal choice means that the complexity of the framework can be scaled down without significant degradation in performance. In essence, users can tailor the framework to their specific needs, allowing for a more adaptable application of our method without sacrificing quality. **Table 1: The effectiveness of each modality. Ablation study on the effectiveness of each modality on 3D object detection and semantic segmentation tasks**. | Image Modality | Text Modality | ScanNetV2 $AP_{25}$ | ScanNetV2 $AP_{50}$ | S3DIS $mIoU$ | S3DIS $mAcc$ | |----------------|---------------|---------------------|---------------------|--------------|--------------| | | | 62.1 | 37.9 | 61.1 | 67.2 | | ✔️ | | 65.7 | 44.6 | 69.0 | 75.4 | | | ✔️ | 64.8 | 43.2 | 68.4 | 75.2 | | ✔️ | ✔️ | **66.3** | **45.5** | **71.1** | **77.5** |

Reviewer hzsQ2023-08-16

Reply to the rebuttal

Most of my concerns are well addressed after reading the rebuttal. I have changed my Rating to Weak Accept.

Authorsrebuttal2023-08-17

Thanks for the revised score

We sincerely appreciate your careful review and consideration of our rebuttal. It's encouraging to hear that we have been able to address most of your concerns, leading to an updated rating.

Reviewer MtnQ2023-08-17

First, I would like to thank the authors for the thorough rebuttal and for providing numbers for numerous experiments. From these new experiments, readers are able to assess the improvement introduced by the method. Regarding the mask guidance, thanks for including this in the final version of the paper. I agree with the authors that these methods cannot be easily transferred to multiple object images. Regarding novelty, my concerns still remain. Combining several foundation models does not seem a big contribution when those have been used separately before. Also, ULIP uses images and text for pre-training, even if it was at object level. The proposed method uses the same ideas at object level. However, since many of my concerns regarding evaluation have been addressed I will increase my score.

Authorsrebuttal2023-08-17

Thanks for the revised score and addressing concerns regarding novelty

Thank you for your thoughtful review and for recognizing the efforts we've made in addressing your previous concerns about evaluation. We're pleased that the additional experiments and the inclusion of mask guidance have helped clarify certain aspects of our work. Regarding the novelty, we understand your reservations about combining several foundation models. However, we believe that the integration of these models in our context is not just a mere combination. In addition to utilizing foundational visual encoders and text encoders found in previous methods, our approach uniquely leverages image segmentation foundation models (such as SAM), image captioning foundation models (like BLIPV2 and Tag2Text), and CHATGPT. To the best of our knowledge, our method represents the first exploration of these foundation models in the realm of 3D scene understanding. Regarding the similarity to ULIP, it is important to highlight some key differences that set our method apart. While the training of ULIP relies on the annotation of 3D objects to extract corresponding text features, this approach is not applicable within 3D scene datasets. Our method uniquely overcomes this limitation by employing image-to-text foundation models to generate relevant text descriptions, further refined using CHATGPT. Moreover, we acknowledge that our method, like ULIP, utilizes both text and image during the pre-training stage. However, it's essential to recognize the distinctions between those two methods. ULIP's instance-level contrastive approach serves well for tasks like instance discrimination but falls short for dense prediction tasks such as object detection and semantic segmentation. This limitation has been observed in various works across 2D and 3D domains [1, 2, 3, 9, 53]. There exists a non-negligible gap between instance-level pre-training as ULIP and our proposed scene-level dense pre-training method. The novelty of our work lies in the development of self-supervised pre-training that is tailored and aligned with 3D object detection and segmentation. This alignment is achieved through the strategic integration of foundation models from both image and text modalities, forming a coherent framework that transcends previous limitations in the field. We sincerely appreciate your feedback and the increased score. Your insights have been instrumental in refining our work, and we're grateful for your constructive engagement with our research. [1] Xie, Zhenda, et al. "Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [2] Wang, Xinlong, et al. "Dense contrastive learning for self-supervised visual pre-training." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021. [3] Liu, Yueh-Cheng, et al. "Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining." arXiv preprint arXiv:2104.04687 (2021).Liu, Yueh-Cheng, et al. "Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining." arXiv preprint arXiv:2104.04687 (2021).

Reviewer E1Dc2023-08-19

I have read your response. Thank you for the segmentation results. I will keep my original score(6).

Reviewer V9nH2023-08-19

Thanks for the rebuttal

The reviewer has carefully read the authors' rebuttal and appreciates the experiments, and would like to keep the original rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC