NeuroGF: A Neural Representation for Fast Geodesic Distance and Path Queries

Geodesics are essential in many geometry processing applications. However, traditional algorithms for computing geodesic distances and paths on 3D mesh models are often inefficient and slow. This makes them impractical for scenarios that require extensive querying of arbitrary point-to-point geodesics. Although neural implicit representations have emerged as a popular way of representing 3D shape geometries, there is still no research on representing geodesics with deep implicit functions. To bridge this gap, this paper presents the first attempt to represent geodesics on 3D mesh models using neural implicit functions. Specifically, we introduce neural geodesic fields (NeuroGFs), which are learned to represent the all-pairs geodesics of a given mesh. By using NeuroGFs, we can efficiently and accurately answer queries of arbitrary point-to-point geodesic distances and paths, overcoming the limitations of traditional algorithms. Evaluations on common 3D models show that NeuroGFs exhibit exceptional performance in solving the single-source all-destination (SSAD) and point-to-point geodesics, and achieve high accuracy consistently. Besides, NeuroGFs also offer the unique advantage of encoding both 3D geometry and geodesics in a unified representation. Moreover, we further extend generalizable learning frameworks of NeuroGFs by adding shape feature encoders, which also show satisfactory performances for unseen shapes and categories. Code is made available at https://github.com/keeganhk/NeuroGF/tree/master.

Paper

References (50)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Ujtd4/10 · confidence 5/52023-07-03

Summary

The paper proposes a neural implicit representation of a 3d surface that enables querying for geodesic lengths and paths between points on the surface. A neural network is overfitted to one given surface. The input query points are embedded into high dimensional features whose euclidean distance represents the desired geodesic distance. Another branch learns to receive the high dimensional features to deform a straight line to the geodesic between the two points. A final branch learns to represent the SDF of the surface. The network is trained via strong supervision w.r.t given geodesics computed on the mesh. The results are shown to produce high accuracy when overfitted to well-known graphics models.

Strengths

The idea of overfitting to represent a geodesic paths can be quite useful, especially considering the ability to parallelize many queries through the network as one batch.

Weaknesses

I find the work to provide a good technical solution to the given problem, however I cannot champion the paper: The technical contribution is quite limited. Beyond the core idea of overfitting to geodesics of a given shape, the rest of the approach is quite straightforward. As such this feels more as an application than a NeurIPS paper.

Questions

1. Why use the SDF at all? why not provide the first 2 branches + the original mesh as the full representation? if the SDF is completely decoupled, what good does it bring. 2. is there any guaranteed that the geodesic paths actually lie on the isosurface of the SDF? if they are not, how is the difference reconciled?

Rating

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

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

2 fair

Limitations

Yes

Reviewer nwTQ6/10 · confidence 3/52023-07-04

Summary

The authors propose a solution to the estimation of the geodesic distance between a pair of points (source, target) on a mesh. The proposed approach relies on an implicit field. In particular, the implicit field learns/memorizes the distance between each pair of points. The authors sample a subset of mesh vertices and pre-compute the geodesic distance between them. These points are later used to optimize the network weights (implicit field). Each point is converted into a feature vector by a shared MLP. The difference between source and target features is used to predict the geodesic distance between the two points. On top of this, a second branch predicts the implicit field representing the shape itself (using the shared MLP features as input). Finally, a third branch using the same features, as concatenation, predicts the geodesic path connecting the source to the target as a set of points. The authors demonstrate the effectiveness of the paper on a handful of meshes, showing numerical and qualitative results, while comparing with other approaches.

Strengths

The paper addresses an important problem. The manuscript describes the idea intuitively, clearly, and in a well-structured manner. The solution is interesting, original, and well-crafted. It leverages past approaches to achieve a different goal. Furthermore, the method is incredibly faster than chosen baselines while being accurate - both in terms of geodesic distance, geodesic path, and reconstruction error. This is possible with the use of NN, as it requires a single forward pass, compared to classic methods which require extensive computations.

Weaknesses

The proposed method uses a shared MLP to predict a set of features from 3D points. However, the input is not limited to points belonging to the surface. The same is the output of the geodesic path (one of the branches). This seems to be quite an important limitation which may severely prevent the application of this method. Indeed, the authors introduce an extra loss term 5) to "force" the points on the curve to belong to the surface. Yet there is no guarantee this happens. Furthermore, to apply this technique we have to optimize a neural network, which takes time while the classic methods do not require such expensive pre-processing.

Questions

The solution is interesting and leaves me with some questions, it would be great if the authors can address some of these: - how much time does the pre-processing time require? ie optimizing the network + sampling points + estimating distances for training - the current formulation for the path relies on AtlasNet/Foldingnet, have the authors thought about using a different approach that cannot predict points off the surface? For example, the use of a transformer that predicts the index of the next point on the path. - how well does the method scale with very large meshes? (#V > 1M) - how robust is the method on sparse meshes where very few vertices can be used to define geodesic distances? For example, on a cube where the only vertices are at the corners, and the query points are pairs on the faces (assuming this is possible). I thank the authors for the time spent addressing any of these concerns.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There are no negative social impacts of this work. The authors discussed the limitation of this work, although they did not detail properly the time requirement of the method (such as pre-processing time). I would recommend the authors add these details either in the main paper or in the supplementary. Furthermore, it would be great if the authors could evaluate the approach on a very diverse set of meshes, such as a cube and a super-dense version of "dragon".

Reviewer yBGS3/10 · confidence 5/52023-07-05

Summary

This paper presents a framework to effeciently compute pairwise geodestic distances and shortest geodesic paths on the surface of a given mesh. This is accomplished by overfititng a neural learning model to a given mesh to regress the distances, paths and a SDF to reconstruct the surface. The geodesic distances/paths are learned in an implicit manner given pairs of points on the surface - specifically, points are sampled along the line between the points and the corresponding points on the geodesic path are predicted by a neural network.

Strengths

The problem being tackled, namely of effecient geodesic computations is important for the graphics community. The method is simple and easy to understand and reproduce.

Weaknesses

1. In my opinion, the overfitting setting weakens the contribution significantly, especially considering that GraphWalks [Potamias et al] performs this task with a dataset and can predict even if with lower accuracy, on unseen test shapes. This work can be extended to learn on datasets, for eg simply by learning a global shape descriptor and appending it to the existing learned point features. 2. Since its in the overfitting sitting, the results are less impressive. Neural networks are known for the ability to interpolate between seen data - i.e, if the model is trained to predict geodesics between enough number of pairs, its not too difficult to then get a good performance on the remaining pairs at inference. 3. Additionally, since it is in the overfitting setting and thus needs to be trained for every new mesh, the training time itself should be considered as part of the time taken to compute the geodesics. The training time for each mesh is not mentioned in the paper. 4. The network takes in point coordinates (via FC layers) and thus the neighborhood of a point is not encoded in any way - a better representation would be to encode as well the immediate neighborhood of the given pair of points. However, this has not been tried in the paper. 5. The role of the SDF is simply as a byproduct, as noted by the authors in the supplemental. Table 3 shows that adding the SDF only provides a small improvement.

Questions

1. The computation of the Distrubution constraint (eq 11) is not clear - the text says that a separate network is trained to predict the SDF of the generated curve poitns P_m, and that it is trained in advance and kept fixed. However, P_m itself is computed during training (from the line sampels), so how is the network trained in advance?

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

1 poor

Limitations

As discussed in the weakness section, the primary limitation is that the method has to be trained for every given mesh, thus I'd consider the training time as part of the geodesic computation time. I'd suggest the authors to extend the method to learn on datasets with a global shape descriptor to make it more useful for the community. In my opinion, the overfitting setting weakens the result and is not as useful. The work would be very useful if it can be used to perform online computations on unseen shapes.

Reviewer oZNk4/10 · confidence 4/52023-07-06

Summary

This paper develops a neural network architecture to estimate the geodesic distances and shortest geodesics between query points on a given 2D surface. It also provides a signed-distance function field evaluated at the given query points. The architecture consists of a set of FC layers followed by Pointwise MLPs. The learning process is supervised. That is, the output of the network is compared with the ground truth distances and geodesic paths. The paper demonstrates these ideas for a number of 3D surfaces and compares the results with two previously published approaches. The results are found to be more accurate and take less computing time.

Strengths

There are many techniques in shape analysis of 3D objects that require computing geodesics or geodesic distances between arbitrary points. This paper develops a learning based solution that provides a slightly more accurate estimates and a somewhat faster pace.

Weaknesses

In my opinion this problem is not that challenging. The paper uses about 20K points on a surface and generates ground truth quantities (geodesic distances, geodesics, etc) from them. This seems to me like a very dense sampling. If on a 2D surface we have pairwise distances between 20K (or some such number) points, then the task of finding distances between remaining arbitrary pairs does not seem that challenging. Indeed, the errors across different methods are not that different – mostly within a one percentage point or so. In terms of regression, one is learning a function f: S x S --> R_+ using millions of data points for the geodesic distance estimation. How about training the algorithm it only hundreds or even thousands of paired distances.

Questions

The conclusion states that the paper learn “neural implicit representations for 3D surface geodesics”. I do not quite understand what this means. Is there a mathematically precise statement that can replace this ambiguous claim.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good 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

3 good

Contribution

2 fair

Limitations

As the authors state, there is no guarantee that the shortest path lies on the actual surface. This seems like a big limitation. How much error does the post-processing introduces in this?

Reviewer Racd5/10 · confidence 4/52023-07-10

Summary

This paper proposed to employ neural geodesic fields to implicitly represent (1) geodesic distance, (2) signed distance filed, (3) shortest geodesic path query using the overfit-paradigm. Experiments are conducted to demonstrate the effectiveness of the proposal.

Strengths

• The writing of this paper is clear and easy to follow. • The extension to use implicit neural representation to represent geodesic distance and efficiency query is novel. It is a simple and yet seems effective approach under the overfitting paradigms for geodesic implicit representation.

Weaknesses

1. The diversity of the mesh shape used in the experiments is limited. No realistic real-world mesh is experimented. The original scale of mesh shape in the experiments is not diverse enough: e.g. from small toy scale to large skyscraper scale mesh to validate the proposal in a more realistic setting.

Questions

1. Page 4, section 3.2, any ablation study on input points embedding? Any comparison regarding FC layers vs position embeddings for point coordinates? Usually position embedding might work better for avoiding over-smoothing of the input signals for complicated shapes. It might worth exploring. 2. Section 3.3, any ablation studies on the learning objectives? Why L1? How about others? 3. Table 1, please annotate the best and worst performing entries in comparison. 4. Section 4.3, more ablation analysis is needed. For example, more conclusive analysis is needed for Table 3.

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

3 good

Contribution

2 fair

Limitations

NA

Reviewer nwTQ2023-08-13

Thank you for taking the time respond to my concern in well a structered manner and the additional evaluation. ### W1 As the shortest path may not belong to the mesh, you are suggesting to perform a local projection. This may solve the problem, or introduce artifacts in concave regions. Either way, the numerical results suggest the error (10^-2) is relatively high compared to what I would agree to be accurate predictions. ### Q1 Thank you for the comprehensive response. Overall a single query may take a long time, even hours, (preprocessing of 21 hours from Table S1 + 119 minutes), this time being amortized with multiple queries. While for a mesh-based approach, each query is computed at runtime. A plot showing #query vs total training time would show the benefits of this approach. I struggle to see where this approach (for SSAD or MSAD) would be applicable, can authors cite any applications where a 10^-2 error in the geodesic error would be acceptable? ### Q3 Thank you for the additional evaluation. ### Q4 Thank you for your response and additional evaluation. I understand with sparse GT samples the accuracy degrades, this is peraphs a limitation worth mentioning in the manuscript.

Authorsrebuttal2023-08-15

**Response:** We sincerely appreciate your evaluation of our rebuttal materials and the active discussion. Here, we further respond to your raised questions with more targeted explanations. In terms of downstream application scenarios, it is worth clarifying that the geodesic representation accuracy achieved by our approach is adequate for supporting a rich variety of geometry processing and modeling tasks. Note that the popular classical geodesic algorithms (e.g., *heat method* (HM) and *fast marching method* (FMM)) typically produce results with higher than 1\% errors (sometimes even much higher for anisotropic meshes). Hence, to answer your question about "*applications where a $10^{-2}$ error in the geodesic error would be acceptable*", basically many papers where (original implementations or variations of) HM and FMM are used to produce geodesic information can be taken as examples. As listed in the following, [A, B, C] involve HM-based and FMM-based approaches for geodesic computation, and [D] simply resorts to constructing k-NN graphs followed by the Floyd’s shortest path algorithm, which leads to even higher errors. Also, it can be used for supporting various interactive geometry processing and modeling tasks, such as texture transfer, decal placement, remeshing, smoothing, splines on surfaces, as reported in [E], in which the authors adopted a low-dimensional Euclidean embedding for computing geodesics with approximation error around 5\%. -- [A] A. Poulenard, et al., "Multi-Directional Geodesic Neural Networks via Equivariant Convolution," in ACM TOG, 2018. -- [B] B. L. Bhatnagar, et al., "Multi-Garment Net: Learning to Dress 3D People from Images," in Proc. ICCV, 2019. -- [C] R. Wiersma, et al., "CNNs on Surfaces using Rotation-Equivariant Features," in ACM TOG, 2020. -- [D] Z. Li, et al., "Geodesic Self-Attention for 3D Point Clouds," in Proc. NeurIPS, 2022. -- [E] D. Panozzo, et al., "Weighted Averages on Surfaces," in ACM TOG, 2013. Moreover, we would like to further explain the actual application scenarios of our approach in an overfitting paradigm. To facilitate the intuitive understanding of what we say "interactive applications with rapid and extensive geodesics querying", here we illustrate a very specific case. In the practical industrial scenarios for game/movie/animation, many important 3D asset models such as characters, animals, or other types of objects, need to be repeatedly manipulated, with extensive querying of geodesics at each time. In this case, it can be highly inefficient to apply conventional computational algorithms for many times, and it is even more impractical to store all pre-computed geodesics for every pair of source-target points due to extremely huge memory cost. For this situation, our NeuroGFs can serve as a good choice. Intuitively, once the offline training (i.e., overfitting) process is completed, the neural network model can be viewed as "an attribute" of the target mesh model, just like other attributes like texture. This trained neural network model is quite compact (around 1MB), fast to online query, and can be permanently stored together with the original mesh model as "a compression of its complete geodesics information". Besides, as for your comment that suggests showing the plot of \#query vs. total training time, we remind that the offline training time cost is not influenced by the number of online queries, because in our implementation we aim to learn complete geodesics of the whole mesh model, regardless of how many queries the users may specify. This setting is reasonable according to the preceding paragraph explaining our suitable application scenarios. More importantly, even if in some cases training time cost is really the most critical consideration factor, one can have multiple choices to deal with the speed-accuracy trade-off: -- ***(a) Preparing sparser ground-truth pairs.*** As reported in Table S1 of the supplementary material as well as Table R3 of the uploaded PDF file, one can choose to generate sparser ground-truth pairs to significantly reduce the time cost of pre-processing procedures, and the resulting geodesic representation accuracy can still maintain satisfactory. -- ***(b) Early stopping of the training process.*** As can be found in our rebuttal response to your [Q1], the fast convergence speed of our NeuroGFs allows us to flexibly control the training time cost while maintaining highly competitive performances. -- ***(c) Adopting generalizable NeuroGF learning frameworks.*** As designed and comprehensively evaluated in our rebuttal (Figure R1 and Table R4 of the uploaded PDF file), one can directly use the generalizable version of our NeuroGF representation framework. After training, the generalizable NeuroGFs can be directly applied to process unseen shapes and categories while achieving satisfactory performances.

Authorsrebuttal2023-08-16

Looking forward to hearing from you

Dear Reviewers **Ujtd**, **yBGS**, **oZNk**, and **Racd**, Thank you for taking the time to review our submission and providing us with constructive comments. We would like to know if our responses adequately addressed your earlier concerns. Additionally, if you have any further concerns or suggestions, we would be more than happy to address and discuss them to enhance the quality of the paper. We eagerly await your response and look forward to hearing from you. Best regards, The authors

Reviewer Racd2023-08-20

Thank you for taking time to respond to my review. Please keep improving the paper. I will keep my original ratings for now.

Authorsrebuttal2023-08-21

Dear Reviewer Racd Thanks very much for your time and efforts in evaluating our rebuttal contents. We will supplement these newly-added discussions and experimental results into the final version to further improve the quality of our paper. Best regards, Authors

Authorsrebuttal2023-08-21

Looking forward to receiving your feedback

Dear **Reviewer oZNk** Thank you for taking the time to review our submission. As the discussion phase between the reviewers and authors is coming to the end, may we know wether there are still unsolved concerns from you? We are pleased to address them. Looking forward to receiving your feedback. Best regards, The authors

Authorsrebuttal2023-08-21

Looking forward to receiving your feedback

Dear **Reviewer Ujtd** Thank you for taking the time to review our submission. As the discussion phase between the reviewers and authors is coming to the end, may we know whether there are still unsolved concerns from you? We are pleased to address them. Looking forward to receiving your feedback. Best regards, The authors

Authorsrebuttal2023-08-21

Dear **Reviewer yBGS** Thank you for taking the time to review our submission. As the discussion phase between the reviewers and authors is coming to the end, may we know whether there are still unsolved concerns from you? We are pleased to address them. Looking forward to receiving your feedback. Best regards, The authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC