Zero-Shot Scene Reconstruction from Single Images with Deep Prior Assembly

Large language and vision models have been leading a revolution in visual computing. By greatly scaling up sizes of data and model parameters, the large models learn deep priors which lead to remarkable performance in various tasks. In this work, we present deep prior assembly, a novel framework that assembles diverse deep priors from large models for scene reconstruction from single images in a zero-shot manner. We show that this challenging task can be done without extra knowledge but just simply generalizing one deep prior in one sub-task. To this end, we introduce novel methods related to poses, scales, and occlusion parsing which are keys to enable deep priors to work together in a robust way. Deep prior assembly does not require any 3D or 2D data-driven training in the task and demonstrates superior performance in generalizing priors to open-world scenes. We conduct evaluations on various datasets, and report analysis, numerical and visual comparisons with the latest methods to show our superiority. Project page: https://junshengzhou.github.io/DeepPriorAssembly.

Paper

References (79)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer veoH5/10 · confidence 3/52024-07-01

Summary

This paper introduces a framework called "Deep Prior Assembly," which combines various deep priors from large models to reconstruct scenes from single images in a zero-shot manner. By breaking down the task into multiple sub-tasks and assigning an expert large model to handle each one, the method demonstrates the ability to reconstruct diverse objects and plausible layouts in open-world scenes without additional data-driven training.

Strengths

The methodology effectively integrates various deep priors from different large models, enhancing the robustness of the reconstruction task and doesn’t rely on 3D or 2D data-driven training, reducing dependency on specific datasets. It introduces new optimization methods for pose, scale, and occlusion parsing, improving the collaborative effect of deep priors.

Weaknesses

1. One obvious drawback of this methodology is the huge memory cost caused by assembling many pre-trained modules if we aim to apply this method on mobile edge device. Can the authors think out of a way to mitigate this problem? 2. The reconstruction quality of this framework relies mainly on the capacity of single-image reconstruction model, e.g. Shap-E in this paper. However, according to practical experience, Shap-E is really sensitive to the scale of input image. Is there a universal appropriate scale that fits to all the instances, as mentioned in the paper the scale is set to 6? 3. The methodology tries to fit the 3D proposal into the scene by aligning them with the depth estimated by Omnidata. According to the ablation, 3D matching plays the most important role compared to SD and 2D-matching in the whole framework. So how to ensure the quality of estimated depth? 4. I am curious about how will the methodology perform on the outdoor scenes.

Questions

Please refer to the weaknesses part.

Rating

5

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

Please refer to the weaknesses part.

Reviewer DEJS5/10 · confidence 4/52024-07-11

Summary

This work introduces a system named "Deep Prior Assembly" for zero-shot scene reconstruction from a single image. It breaks down the single-image scene reconstruction task into several steps that can be solved utilizing pretrained large models, such as SAM for segmentation, Shap-E for 3D object generation, and Omnidata for depth estimation. Additionally, an optimization-based approach is also proposed for pose estimation.

Strengths

**1. It makes sense to break down a difficult task into simpler ones, and the outcomes appear promising.** This study puts into practice the concept of breaking down the challenging single-view reconstruction task into several simpler tasks that can be addressed using off-the-shield pretrained models. **2. Experiments and ablation studies are thorough.** Extensive quantitative and qualitative evaluations are provided in the paper, demonstrating the effectiveness of the proposed pipeline.

Weaknesses

**1. Technical contribution of this work is limited.** The work seems to be more of an engineering trial that combines pretrained models to build a reconstruction system, rather than an insightful research endeavor. The authors highlight that they are exploring "deep prior" from large pretrained models, but the priors are not well integrated and instead function independently as separate modules. Additionally, naively combining many large models could lead to high computational demand and potential error accumulation. For example, the proposed system will run a diffusion generative model several times for each object in a scene, and includes a 9.2-second optimization for pose estimation of each object. **2. Missing state-of-the-art baseline methods.** Although there are thorough evaluation results provided in the paper, all the baseline methods are proposed before 2021. Has there been any new progress in single-view scene reconstruction after 2021? For example, ScenePrior [1] published in CVPR'23 introduces a conditional autoregressive generative method for single-view reconstruction. --- [1] Learning 3D Scene Priors with 2D Supervision. Nie et al. CVPR 2023.

Questions

**1. What time does each stage of the proposed method take? And what time do the baseline methods take?** **2. What's the relation between the entire running time of reconstruction and the number of objects in the scene? And how about the baseline methods?** **3. Have you considered using methods like 3D object detection for layout estimation? It could potentially offer faster and more accurate results than the proposed depth-based method.**

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

**1. A naive combination of pretrained large models might cause error accumulation.** **2. The proposed method has a much longer running time compared to baseline methods.**

Reviewer 67ke5/10 · confidence 3/52024-07-13

Summary

The method uses leverages multiple off-the-shelf models to parse a scene represented by a single image into 3D assets in their respective layout. Concretely they use a segmentation model to locate objects, then a diffusion model to enhance the image quality, use Shape-E to generate 3D proposals, and estimate the layout use depth estimation and point cloud matching. The method outperforms existing methods.

Strengths

**Originality:** I think the originality is relatively low as all of the components in the system are off-the-shelf model and they are combined in a relatively straight-forward manner. **Quality:** The results look good both quantitatively and qualitatively. I am less familiar with the task this work is trying to solve, so I cannot speak to how significant the quantitative improvements are or the strength of the baselines. **Clarity:** Overall the figures illustrate the method well and the qualitative examples demonstrate the results. **Significance:** I cannot speak to the significance of the method or results as I am not that familiar with this task or the relevant literature. I think it would help to spend more time in the introduction discussing why this task is important and what are the applications. VR/AR is mentioned in passing, but more concretely anchoring the task to a specific application would help motivate the work.

Weaknesses

The method relies on the quality of foundational models to perform well. The comparison to other baselines is not apples to apples since the other methods are not trained on the large datasets leveraged by the foundational model. This isn't necessarily a flaw in this work, since leveraging additional data could be seen as a potential advantage, but it does create an unequal comparison. Like I mentioned in the strengths section, I think it's unclear what the motivation for the paper is. A little further discussion to anchor this work to an application would be helpful since the method novelty is not particularly high.

Questions

Does this work for other types of scenes such as outdoor? Or is restricted to indoor scenes with furniture? **Suggestions:** Figure 2 has a lot of elements and is hard to parse. Maybe simplifying the figure to illustrate the main idea would help.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes limitations are discussed.

Reviewer osDy5/10 · confidence 4/52024-07-16

Summary

A multistage pipeline for single image 3D reconstruction is proposed, leveraging multiple off-the-shelf models. To begin, SAM is used to segment and decompose the input image. Stable diffusion is then leveraged to complete instance segments with potentially missing information, and failures of this process are filtered out by CLIP. Finally, Shap-E is applied to generate 3D models, which are then registered to the image for a final 3D reconstruction.

Strengths

The proposed technique achieves strong zero-shot performance relative to baseline methods despite not training on similar data.

Weaknesses

The method seems overly complicated and unlikely to be robust. * The argument about heuristic selection of depth shift seems unconvincing; in practice, the correct depth shift from a scale and shift invariant monodepth estimator can vary widely between multiple images even in the same scene. Why not use multiple pairs of images to compute the appropriate depth shift, or metrically ground the depth shift as in RealmDreamer? * In order to go from a SAM instance to mask, a very complex pipeline is proposed. Essentially, it consists of inpainting with SD, followed by filtering with CLIP, followed by shape estimation with Shap-E, followed by a likely unstable alignment procedure. This seems error-prone. Why not just train an LRM-like model that accepts potentially off-center and partially occluded instance images and produces 3D objects aligned to the input camera’s location? This seems not too difficult to train, there are 10M+ object datasets publicly available to achieve it, and it would certainly be more robust than the proposed pipeline.

Questions

Since the method consists of chaining several foundation models together, it should be possible to show "zero-shot" performance on scenes not limited to the simple synthetic or indoor room scenes shown here. Can the proposed method work on more complex scenes, such as real images, outdoor scenes or scenes containing animals or people?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes, there is a thorough limitations discussion.

Area Chair xuWB2024-08-12

Rebuttal

Dear reviewer, Please read the author rebuttal and the other reviews and post a comment as to how your opinion has or has not changed and why.

Reviewer osDy2024-08-13

Thanks for the thoughtful rebuttal. I'll change my score to borderline accept. I'm still unconvinced about the pipeline setup. You might consider https://zixuanh.com/projects/zeroshape.html (in particular how they handle their training data) as an alternative and competing approach which would be likely to have strong performance if trained in the specified way. The results on in the wild images are impressive! I would definitely suggest featuring these (and more + other examples) more prominently in a revised version of the manuscript, since the existing datasets are rather monotonous and supervised methods would probably perform well, so the "zero-shot" aspect is not really emphasized.

Authorsrebuttal2024-08-13

Response to Reviewer osDy

Dear Reviewer osDy, Many thanks for all the helpful comments and positive assessment. We will include the new experiment results and more reconstruction sampls in the revised version of the manuscript. Following your suggestions, we consider training a reconstruction model similar to ZeroShape for predicting complete shapes as a future work of DeepPriorAssembly. We will include the insights to Sec.5 of the revised paper and will conduct experiments to explore the effectiveness of this alternative. We really appreciate you for upgrading the score. Best regards, Authors

Authorsrebuttal2024-08-08

We will be happy to take any questions

Dear reviewers, We appreciate your comments and expertise. Please let us know if there is anything we can clarify further. We would be happy to take this opportunity to discuss with you. Thanks, The authors

Authorsrebuttal2024-08-11

Dear reviewers, As the reviewer-author discussion period is about to end, we are looking forward to your feedback on our rebuttal. Please let us know if our responses address your concerns. We would be glad to make any further explanation and clarification. Thanks, The authors

Reviewer veoH2024-08-10

Thanks for your explanation. 1. As this paper dedicated to assemble different priors from large foundation models, I think it's more convincing to show the robustness of the pipeline by substitute different parts with similar foundation models. 2. Moreover, I think it's inappropriate to claim your method's superior performance on zero-shot manner, because this capacity actually comes from the massive training data of Shap-E. Also, as we can observe from the figure A in the rebuttal PDF, the method fails to perform well on outdoor scene, which is limited by Shap-E. Rationally, I would like to categorize this method as a straight-forward engineering effort to assemble the foundation models. Hope the authors can optimize the connection and compatibility between different priors.

Authorsrebuttal2024-08-10

Responses to Reviewer veoH

Dear Reviewer veoH, Thank you for your response and the helpful comments. We respond to each of your additional questions below. Please do not hesitate to let us know if you have any additional questions. **Discussion-Q1: Ablation studies on the choice of foundation models.** We fully recognize the importance of evaluating each sub-task by substituting the foundational models with similar alternatives. Actually, we have conducted the ablation study to explore the effectiveness of our chosen solutions in each sub-task by comparing them with the alternatives. The results and analisis are presented in Sec.C.1 and Table 4 of the Appendix. Specifically, we conducted ablations to replace Shap·E with One-2-3-45, replace Open-CLIP with EVA-CLIP and replace Omnidata with MiDaS, observing performance degradation with the alternative models. We further visually compare Shap·E with One-2-3-45 for shape generation in Fig. 8, where the results demonstrate that Shap·E is a more robust solution for generating 3D models from 2D instances. These ablation studies validate the effectiveness of each choice made within our framework. **Discussion-Q2: The claim of zero-shot scene reconstruction.** Actually, the term "zero-shot" signifies that no specific data or data-driven training is required for the task of scene reconstruction from single images. Unlike previous works in this field, such as PanoRecon and Total3D, which require specific image-scene pair data, our approach does not rely on such data. The training data used in Shap·E is only capable of shape reconstruction learning, and it is impossible to train a model for single-view scene reconstruction solely by relying on the Shap·E data. We justify that for the "zero-shot" tasks, using large-scale data from a different domain or task is not prohibited. For example, CLIP models are widely used for zero-shot image classification without requiring specific image-class pair data, though they do require massive amounts of image-text pair data for contrastive learning. Similarly, the data used in Shap·E does not undermine our claim of "zero-shot" single-view scene reconstruction. We will provide further clarification on the tasks and claims to enhance the understanding of DeepPriorAssembly's capabilities and the data used in each foundational model integrated into our framework. **Discussion-Q3: Performance on outdoor scenes.** We would like to justify that none of the previous works can successfully reconstruct outdoor scenes from single images. All prior approaches are trained on indoor scenes and struggle to generalize to real-world outdoor images, which contain out-of-distribution objects such as trees, buildings, and animals. Leveraging powerful large foundational models, DeepPriorAssembly is the first to demonstrate the capability for zero-shot scene reconstruction in these complex real-world outdoor scenes. As illustrated in Fig.A of the rebuttal PDF, DeepPriorAssembly accurately reconstructs scene geometries even in challenging outdoor scenes. However, the texture produced by Shap·E may not be optimal for the outdoor shapes. In the future, we may consider replacing Shap·E with a shape reconstruction method that performs better on outdoor shapes to further enhance the performance of outdoor scene reconstruction. **Discussion-Q4: Optimize the connection and compatibility between different priors.** Thanks for your suggestions. In the revision, we will improve the connection and compatibility among the foundation models by conducting more ablation studies on the choice of each foundation model within our framework, supplementing the analysis presented in Sec.C.1 and Table 4. We are deeply grateful for your invaluable feedback and the time you dedicated to evaluating our work. Your comments and expertise are sincerely appreciated. Please let us know if there is anything we can clarify further. Best regards, Authors

Reviewer veoH2024-08-12

Thanks for your response. I appreciate your effort on ablation studies for different foundation models. I think the ability of "zero-shot" within your method comes from the off-the-shelf foundation models(mainly) and the divide-and-conquer thought. So this should not be a prior advantage over other data-driven methods. In my opinion, the authors should design extensive experiments to demonstrate the advantages of the way of their combination, rather than "combination" itself. Overall, I think the technical contribution of this assembly is still limited, so I will keep my score as 5.

Authorsrebuttal2024-08-12

Response to Reviewer veoH

Dear Reviewer veoH, Thank you for your response and the positive assessment. We really appreciate your expertise and all the invaluable feedback. We response to each of your additional questions below. **Discussion-Q5:Efforts on the combination way of deep priors.** We have invested extensive efforts in exploring the optimal approach to robustly integrating available deep priors for a challenging task, rather than merely combining them. We demonstrate that the naive approach of simply combining several large models fails to solve the challenging task of zero-shot scene reconstruction. Beyond the contribution of firstly proposing to assemble large models and the task decomposition for this challenge task, we offer additional significant technical contributions on solving the critical challenge of making deep priors work together robustly. Specifically, the naive solution involves segmenting the input scene images and then generating 3D objects from the segmented instances. However, the solution fails dramatically due to: (1) the instances are often corrupted by occlusions and low-resolution, leading to failures for reconstructing complete and high-quality 3D objects, and (2) none of the existing techniques are capable of recovering the scene layout for 3D objects. To address these challenges and robustly assemble deep priors for zero-shot 3D scene reconstruction, we propose two significant technical contributions on improving the quality of 3D instances and accurately recovering the scene layouts. (1) To improve the robustness of the framework and overcome the challenges of occlusions and low-solution in segmented instances, we novelly introduce the StableDiffusion model to enhance and inpaint the instance images, followed by CLIP models to filter out the poor-quality samples and select the ones matching the instance most. The novel designs on task decomposition and introducing suitable deep priors are the key contributors for achieving accurate and high-quality geometries and appearances of the generated shapes. (2) For the challenging task of recovering the layout of the scene containing the reconstructed shapes, we propose a novel approach on optimizing the location, orientation and size of instances by matching them with both 2D and 3D supervisions. The supervisions are novelly introduced from the estimated segmentation masks and the predicted depths, which are also obtained by the assembled deep priors. Moreover, a RANSAC-like solution is proposed to further improve the robustness of the pose/scale optimization. This approach is the link among deep priors and plays the key role in robustly assembling deep priors for the final target of zero-shot scene reconstruction. Following your suggestions, we will conduct more experiments on the combination way of deep priors in our revision. **Discussion-Q6:Technical contributions.** Our technical contributions lie in the novel task decomposition, the deep priors chosen and their assembly, and the proposed RANSAC-like pose/scale optimization for recovering scene layouts. We summarize the main contributions as follows. 1. We propose the first framework which assembles diverse deep priors from large models together for the extremely difficult task of reconstructing scenes from single images in a zero-shot manner. 2. To improve the robustness of the framework and overcome the challenges in this task (e.g. occlusions and low-solution of instances), we novelly utilizes the StableDiffusion model for image enhancement and inpainting, combined with the CLIP model to filter out poor-quality samples. 3. We introduce a novel approach on optimizing the location, orientation and size of instances by matching them with both 2D and 3D supervisions. Moreover, a RANSAC-like solution is proposed to further improve the robustness of the pose/scale optimization. The approach is the link among deep priors and plays the key role in robustly assembling deep priors for the final target of zero-shot scene reconstruction. Only with our designs on the task decomposition, deep prior chosen and our novel approach on RANSAC-like pose/scale optimization through both 2D and 3D matching to recover the scene layout, can the assembly of deep priors from large models effectively succeed in the extremely challenging task of zero-shot scene reconstruction. We are deeply grateful for your invaluable feedback and the time you dedicated to evaluating our work. Your comments and expertise are sincerely appreciated. Please let us know if there is anything we can clarify further. Best regards, Authors

Reviewer 67ke2024-08-12

Response to Rebuttal

I thank the authors for their response. After reading other reviews and other responses to reviewers I will maintain my rating of borderline, 5. I think the technical contribution is limited as foundational models are being fused together and the techniques for combining don't seem particularly general. I think the motivation for matching assets to images of scenes is rather limited in its current form. In robotics, meshes with 6 DoF are estimated from images for grasping, but this task is a bit far from those in robotics. For the positives the paper improves over current methods, the figures and qualitative results are well done, and the writing and organization is clear. Therefore I maintain my rating of 5.

Authorsrebuttal2024-08-12

Response to Reviewer 67ke

Dear Reviewer 67ke, Many thanks for the positive assessment. We really appreciate your expertise and all the invaluable feedback. Our method focuses on how to robustly join available deep priors for a challenging task. For the technical contributions, we make additional clarifications below. We justify that the naive approach of simply introducing several large models fails to solve the challenging task of zero-shot scene reconstruction. Beyond the contribution of firstly proposing to assemble large models and the task decomposition for this challenge task, we offer additional significant technical contributions on solving the critical challenge of making deep priors work together robustly. Specifically, the naive solution involves segmenting the input scene images and then generating 3D objects from the segmented instances. However, the solution fails dramatically due to: (1) the instances are often corrupted by occlusions and low-resolution, leading to failures for reconstructing complete and high-quality 3D objects, and (2) none of the existing techniques are capable of recovering the scene layout for 3D objects. To address these challenges and robustly assemble deep priors for zero-shot 3D scene reconstruction, we propose two significant technical contributions on improving the quality of 3D instances and accurately recovering the scene layouts. (1) To improve the robustness of the framework and overcome the challenges of occlusions and low-solution in segmented instances, we novelly introduce the StableDiffusion model to enhance and inpaint the instance images, followed by CLIP models to filter out the bad samples and select the ones matching the instance most. The novel designs on task decomposition and introducing suitable deep priors are the key contributors for achieving accurate and high-quality geometries and appearances of the generated shapes. (2) For the challenging task of recovering the layout of the scene containing the reconstructed shapes, we propose a novel approach on optimizing the location, orientation and size of instances by matching them with both 2D and 3D supervisions. The supervisions are novelly introduced from the estimated segmentation masks and the predicted depths, which are also obtained by the assembled deep priors. Moreover, a RANSAC-like solution is proposed to further improve the robustness of the pose/scale optimization. This approach is the link among deep priors and plays the key role in robustly assembling deep priors for the final target of zero-shot scene reconstruction. Our technical contributions can be summarized as follows. 1. We propose the first framework which assembles diverse deep priors from large models together for the extremely difficult task of reconstructing scenes from single images in a zero-shot manner. 2. To improve the robustness of the framework and overcome the challenges in this task (e.g. occlusions and low-solution of instances), we novelly utilizes the StableDiffusion model for image enhancement and inpainting, combined with the CLIP model to filter out poor-quality samples. 3. We introduce a novel approach on optimizing the location, orientation and size of instances by matching them with both 2D and 3D supervisions. Moreover, a RANSAC-like solution is proposed to further improve the robustness of the pose/scale optimization. The approach is the link among deep priors and plays the key role in robustly assembling deep priors for the final target of zero-shot scene reconstruction. Only with our designs on the task decomposition, deep prior chosen and our novel approach on RANSAC-like pose/scale optimization through both 2D and 3D matching to recover the scene layout, can the assembly of deep priors from large models effectively succeed in the extremely challenging task of zero-shot scene reconstruction. We are deeply grateful for your invaluable feedback and the time you dedicated to evaluating our work. Your comments and expertise are sincerely appreciated. Please let us know if there is anything we can clarify further. Best regards, Authors

Reviewer DEJS2024-08-12

I appreciate the detailed response to the review and have also read the comments from other reviewers. While the proposed method presents impressive experimental results, it is more like an engineering effort by combining several pretrained modules to build a software pipeline, which is not quite suitable for the NeurIPS research community. Therefore, I maintain my original score of "5: Borderline accept" and would like to participate actively in the subsequent reviewer discussion period.

Authorsrebuttal2024-08-13

Response to Reviewer DEJS

Dear Reviewer DEJS, Thank you for your response and the positive assessment on our rebuttal. We really appreciate your expertise and all the invaluable feedback. The technical contributions of DeepPriorAssembly lie in the the exploration of the robust and optimal approach to robustly integrating available deep priors for a challenging task, rather than simply combining them. We show that simply combining several large models is not enough to solve the difficult task of zero-shot scene reconstruction. While our work is the first to propose assembling large models and decomposing tasks to tackle this challenge, we also make key technical contributions to ensure deep priors work together effectively and robustly. The naive approach involves segmenting scene images and generating 3D objects from these segments, but it fails because (1) occlusions and low resolution often lead to incomplete or poor-quality 3D objects, and (2) existing techniques can't recover the 3D scene layout. To overcome these issues, we propose two main techniques: 1. To enhance the quality of 3D instances and handle occlusions and low resolution, we introduce the StableDiffusion model to refine and inpaint segmented images, followed by CLIP models to filter and select the best-matching instances. These innovations in task decomposition and deep prior integration are crucial for producing accurate and high-quality 3D shapes. 2. For recovering the scene layout, we introduce a novel method to optimize the location, orientation, and size of instances by matching them with both 2D and 3D supervision. These supervisions come from estimated segmentation masks and predicted depths, generated by the assembled deep priors. Additionally, a RANSAC-like solution further improves the robustness of pose and scale optimization. This method effectively links the deep priors and is key to achieving robust zero-shot scene reconstruction. Thank you for your valuable feedback and the time you spent evaluating our work. We truly appreciate your insights. Best regards, Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC