NeurIPS 2020

SDF-SRN: Learning Signed Distance 3D Object Reconstruction from Static Images


Review 1

Summary and Contributions: This paper presents a differentiable rendering framework for implicit surface based on Scene Representation Network [38]. It refines the formuation of [38] by leveraging the distance transformed sihouettes for richer geometric supervision.

Strengths: - An improved differentiable rendering framework for implicit surface is proposed, which can further improve the rendering accuracy of scene representation network [38].

Weaknesses: - The writing of this paper is very misleading. First of all, it claims that it can be trained only using a single viewpoint of the object. In fact, all previous diffrentiable rendering techniques can be trained using a single view of object at training time. However, the reason why multi-view images are used for training in prior works is that single-view images usually lead to ambiguity in the depth direction. The proposed method also suffers from this problem -- it cannot resolve the ambiguity of depth using a single image either. The distrance-transformed silhouette can only provide information on the xy plane - the shape perpendicular to the viewing direction. - I doubt the proposed method can be trained without using any camera information (Line 223, the so called "knowledge of CAD model correspondences"). Without knowing the viewpoint, how is it possible to perform ray marching? How do you know where the ray comes from? - The experiments are not comprehensive and convincing. 1) The comparisons do not seem fair. The performance of DVR is far worse than that in the original DVR paper. Is DVR trained and tested on the same data? What is the code used for evaluation? Is it from the original authors or reimplementation? 2) Though it could be interesting to see how SoftRas performs, it is not very fair to compare SoftRas here as it use a different 3D representation -- mesh. It is well known that mesh representation cannot model arbitrary topology. Thus it is not surprising to see it is outperformed. Since this paper works on implicit surface, it would be more interesting to compare with more state-of-the-art differentiable renderers for implicit surface, i.e. [26],[27],[14], or at least the baseline approach [38]. However, no direct comparisons with these approaches are provided, making it difficult to verify the effectivenss of the proposed approach.

Correctness: Many of the claims are incorrect or overclaimed. Just to name a few. For instance, Line 42-43: "In this paper, we make significant advances on learning dense 3D object reconstruction from single images and silhouettes, without the knowledge of the underlying shape structure or topology." -- This has been already achieved by many of previous works, i.e. [26],[27],[14]. Please refer to the weakness section for more details. Method-wise, the proposed method seems to be able to improve the rendering accuracy. However, the provided experiments are not convincing to verify this point.

Clarity: This work fails to achieve the contributions that claimed in the paper. The entire positioning of this paper is very misleading. Some of the claims are incorrect based on the proposed method. Please refer to the weakness sections for more details.

Relation to Prior Work: This paper is overclaiming its contribution, as elaborated in the weakness section, when discussing its advantages over the previous works. I would highly recommend the writing to be adjusted accordingly in the revisions.

Reproducibility: No

Additional Feedback: It is not clear how the proposed method is able to train network without using any viewpoint information as claimed in the paper. More details are needed to clarify this point. -- post rebuttal -- The authors' rebuttal still cannot convince me that the proposed approach can resolve the ambiguity in the depth direction and thus be able to learn 3D shapes only using a single image for each object during training. This is very important because if the framework cannot resolve the ambiguity in the depth direction, all the contributions and claims of this work do not stand. 1) In the rebuttal, the author claimed that single-view training is possible because they exist at the semantic level and use CMR as a successful example. However, CMR resolves the depth ambiguity using an initial category-specific 3D template, introducing a very strong prior/initial value for the depth. Hence, CMR only needs to learn a plausible deformation of the template without the need to estimate the depth from the ground up. In contrast, the setting of SDF-SRN is totally different in that we have no prior knowledge of the shape but only an image. And thus the success of CMR cannot provide any insight that current framework can be successful. 2) As also pointed out by R4, the 2D SDF loss can only provide supervision on the contour of a one-view visual hull as only the silhouette is used. The ray sampler can only help to sample more efficiently but cannot resolve the ambiguity in the depth direction either. The color loss can be fooled by a flat plane with the exact same texture with the input image. None of these loss terms can resolve the ambiguity of the depth without using multiple views. 3) The authors mentioned in the rebuttal that the ray marching method can learn semantic correspondence, which is very vague. In the presented method, I do not see any part of it that can learn semantic correspondence across objects.


Review 2

Summary and Contributions: The paper introduced a novel formulation to optimize SDF representation from 2D distance transformations, and achieves dense 3D object reconstruction outputs. The paper also proposed a differentiable rendering framework that are suitable for SDF representations. Experimental results on synthetic and real images demonstrate that the proposed approach is significantly better than existing algorithms.

Strengths: The paper is well-written, and contain several professional/beautiful visualizations that can help readers understand the approach. The contributions of combining SRN, SDF prediction, as well as the design of those loss functions are all nicely connected. Experimental results also suggest that the approach achieves great results comparing to existing approaches.

Weaknesses: This is a nicely written paper, but I have some constructive comments, and I hope authors can address those issues in future revisions. - I wonder if the proposed approach will generalize well to objects with extreme poses (such as a chair where the majority of the region is occluded). It seems that under the current formulation, a significant amount of object parts need to be visible. With multi-view supervision, the approach seems to be achieving better performance. How are those viewpoints being chosen in training? - What is the run-time of the proposed approach? I think it may depend on the LSTM steps too. Could authors comment on how those timesteps are determined, and the comment on the output quality improvements along the way? - DVR[33] was not utilizing silhouettes, so I wonder if that is a fair comparison or not. - In real-world applications, the silhouettes are usually not available. Therefore users may need to hand-draw such masks.

Correctness: Yes.

Clarity: Yes, I especially like the diagrams and visualizations of this paper.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback:


Review 3

Summary and Contributions: The paper proposes a method for single-view-3d reconstruction. The method is trained with image, silhouette, and camera-pose ground truth. It leverages a distance transform of the silhouettes to train a set of networks for regressing a 3D signed distance function. A hypernetwork predicts from an image weights of MLPs and an LSTM, which in turn predict occupancy and color given 3D points.

Strengths: + The proposed method allows learning without explicit 3D geometry supervision and only from a single view per object. + The method is well described and the paper clearly written. + The method is evaluated against strong baselines and performs better. + Training on just a single view per object as proposed is interesting for learning from novel categories where 3D CAD models for supervision may not be readily available.

Weaknesses: - The dependence on the camera pose somewhat limits the applicability. It also raises the question how much the proposed method relies on high quality camera poses. Similarly, it relies on high quality silhouette images and it is not quite clear how robust the method would be in practical setups with silhouettes coming from e.g. an instance segmentation algorithm. An experiment with different levels of noise for the camera parameters and less than perfect silhouettes would be helpful to asses the robustness in real-world settings. - I would appreciate a more detailed motivation for the hypernetworks. Other work directly optimized the MLPs or conditioned predictions from the MLPs via latent codes. I would suspect those approaches to be more robust than a network predicting network weights. Further, I expect the hypernetwork setup to work worse for a given memory budget than the above discussed other options. An ablation experiment could address these concerns and provide evidence for eventual benefits of the proposed setups. - For evaluation, the method is trained per category, which is a more restrictive setup than most methods are evaluated on. While there is no explicit explanation for this choice of setup, I suspect that otherwise considerably larger MLPs would be required for predicting sufficiently accurate shapes. Predicting the weights of those larger MLPs may be much harder with the hypernetwork or may not work well at all. Furthermore, only 3 of the 55 ShapeNet categories are considered in the experiments. Is that because these are the largest in ShapeNet and all other categories do not provide sufficiently many shapes for learning a decent model? The experiment in Table 2 suggests that the method should be well suited for decent results especially if few models are available. Previous work for single-view 3d reconstruction has commonly evaluated methods trained on all or at least 13 categories jointly. I would thus very much appreciate experiments for more categories and ideally including joint training.

Correctness: The claims and methods appear correct.

Clarity: The paper is well written and easy to follow.

Relation to Prior Work: Prior related work has been sufficiently discussed.

Reproducibility: Yes

Additional Feedback: Reading other reviewer's comments, the paper again, and comparing to the original DVR paper, I have become more concerned about the DVR comparison. The results shown for DVR in this paper are terrible and do not correspond at all to results in the original DVR paper, which is especially concerning as originally DVR was trained on more ShapeNet categories. The reason seems to be the carefully chosen setup for which multi-view consistency for DVR is removed. While the setup is presented as training with single views, up to 24 views of a shape are used during the ShapeNet training nevertheless. I nevertheless think that the paper proposes some interesting ideas, which to me lift it slightly above the acceptance threshold.


Review 4

Summary and Contributions: This submission proposes a method for 3D-unsupervised single-view 3D reconstruction. Unlike previous methods which relied on “multiview” training setting, this work proposes the method that could work on “multi-view” training setting (without view/instance association) and “single-view” training setting. The techniques that differ from previous methods are that the paper : 1. observes the relationship between 2D-SDF and 3D-SDF (the proposition on line 116), and uses it better to supervise SDF outside the visual hull. The loss can be regarded as to enforce an “order” of signed distance instead of only a binary supervision. 2. incorporate an additional network (h) to learn to sample along the ray (potentially more efficiently than pre-design procedures as in NeRF or DVR). A loss function is proposed to enforce those samples before/after hitting the surface to have more distinctive SDF values. Empirically, the paper presents better results in single-view 3D reconstruction, trained under the single-view setting. Further the work also shows its performance on more realistic dataset PASCAL 3D+.

Strengths: The paper addresses an important problem that in the single-view 3D reconstruction method, existing methods tend to rely on the multiview training setting (multi-view images and poses are available for 3D instances). The single view setting or multiview view setting (without instance correspondence) would potentially be very important, as it’s harder to capture and annotate multi-view setting datasets. The design of 2D-SDF loss (L_{SDF}) appears to provide stronger supervision signals, compared to binary occupancy loss. The ray marching network (network h) could potentially accelerate the ray marching process, which in turns could accelerate train/inference speed. Well-written paper and good presentation.

Weaknesses: 1. The work presents good results (qualitative and quantitative) compared to existing methods SoftRas, DVR and CMR. However, I still found it hard to understand which components contribute to the improvement. The underlying difficulty of the single-view training setting lies in the ambiguity of shape and texture: it could be any camera-aligned surface, where the textures are projected onto the surface. The multi-view training (without view/instance association) is similar to this case, with a bit more information in the dataset that there are some same instances (but not explicitly provided to the network) The 2D-SDF loss enforces the learnt SDF outside the visual hull to have an “order”: at least the value from the processed silhouette image (2D-SDF). It doesn’t seem to enforce within the visual-hull volume, therefore the ambiguity still is there. The ray marching network (h) potentially learns to sample the ray in fewer samples; that’s good. The ray loss makes sense to make samples from two sides of the surface to be distinctive to the other side. However, this relies on the h network learnt to sample at the correct depth. In another word, the two networks are collaborating to find a good surface, where they can achieve the agreement at many solutions. Since we don’t have multiview supervision, this is still ambiguous. RGB loss might not help the situation. You can imagine you can print the image on a flat surface. That would still satisfy the loss. The Eikonal regularizer might help in a more “conservative” prediction, as other papers also explore Laplacian loss and flatness loss (e.g. SoftRas). Maybe more explanation would be helpful for the audience to know what has helped the training. 2. The direct comparison case: DVR is also an implicit function based method. The proposed method has a significant advantage over DVR. What might the reason be? To my understanding, DVR utilizes the binary occupancy (SDF-SRN uses the 2D-SDF loss), DVR also did coarse-to-fine ray sampling (uniform sampling then solving for intersection depth by iterative secant method). SDF-SRN learns a network to perform ray sampling: I expect it to be faster, but would it be helpful in learning correct shape without multi-view info? DVR didn’t apply position encoding or Eikonal regularizer. Would these be important factors? Maybe some explanation would be great. 3. Mentioned in 1, there is possibly ambiguity between 3D sdf function f(.) and ray marching function h(.). Imaging network h has recognized the surface as a flat surface, then the f function will reinforce this knowledge. What info does the two explore to get correct supervision (with no multiview info)? 4. Qualitative results of SDF-SRN show “repetitive” artifacts (in most results, especially in Fig 4, 5, 6), . Any explanations?

Correctness: Please see the weakness session

Clarity: Mostly yes Well written paper

Relation to Prior Work: Quite complete related work, although I would love to see the comparison / discussion as mentioned in the weakness session

Reproducibility: Yes

Additional Feedback: ** After rebuttal ** Thank you for the very detailed rebuttal. After reading the rebuttal and discussion with other reviewers, personally speaking, I still hold the following concerns: 1. On the ambiguity issue, the authors gave some explanation on learning a "semantic" correspondence. It's a bit ambiguous explanation. In my most supportive interpretation, is that, e.g., the method learns the shape prior of a chair (how a chair would look like under various views), as in many other single-view image-to-3D regression system. But those methods either have direct supervision or multi-view self-supervision (results at least better than traditional visual hull). But as mentioned in weakness 1 in my original review, to my understanding, no supervising signal is available to shape the chair in correct shape. To my speculation, it might be the dataset setting (shapenet CAD objects with minimal texture, single category, limited camera views), or network weights are nicely initialized (as in Deep Image Prior, CVPR 18 and Deep Geometric Prior for Surface Reconstruction, CVPR 19). The authors mentioned CMR, but CMR also utilizes keypoint annotations, which is an indirect way to provided cross-view association. More discussion would be crucial to convince the reviewers and future audiences. Please also add the details of network initialization (image encoder and the LSTM). 2. Why the proposed method does so much better than DVR (quite similar loss function design). The authors rule of the possibility of influence of positional encoding and eikonal term. The LSTM ray-marching function seems to be stronger than "naive" ray probing, but the proposed method has a "duplicated" design of f() and h() (i.e. f() also does the job to tell if a point is on surface, and h() specializes on searching on a ray), while DVR marches in ray on by guidance of f(). If f() and h() are learned at the same time (so we can regard it as one black box), why it does so much better another black box of f() alone? The authors indeed pointed out binary occupancy could play an important role. In that sense, would DVR be much better if trained with L_SDF and L_ray? Here some additional ablation and discussion would be important. Also please confirm if DVR is trained under exactly the same setting as the proposed method. For the two reasons above, I'm still leaning to reject. Considering this is a move for single-view training and partially-addressed rebuttal, I would slightly raise the rating to 5: marginally below threshold.