• Submissions
  • Select Your Role : 
  • Author
    • Author
    • Reviewer
  • ICCV2019
    • Conferences
    • AAAI2020
    • AISTATS2020
    • CVPR2020
    • ICCV2019
    • ICML2019
    • NeurIPS2019
  • Zhen Zhang
    • General
    • User Profile
    • Change Email
    • Change Password
    • Link to Account
    • Linked Accounts
      • [email protected] Switch Unlink
    • Delete Account
    • ICCV2019
    • Domain Conflicts
    • Email Chairs
    • Logout

Change Password

Please fill out the form correctly.
Cancel

User Profile

Your profile settings are invalid. Please fill out correctly.
Request account deletion Cancel

Change Email

Please fill out the form correctly.
Cancel
 Print

View Reviews

Paper ID
579
Paper Title
Deep Graphical Feature Learning for the Feature Matching Problem
Reviewer #1

Questions

  • 1. Summary. In 3-5 sentences, describe the key ideas, experiments, and their significance.
    • The paper proposes a graph neural network based method for feature matching. The key idea is to transform 2D coordinates of feature points into local features that can be used for comparison and matching. The key contributions are: a geometric feature net that takes Nx2 point coordinates and transforms them to Nx512 features, a compositional message parsing algorithm (a soft-version of MPNN) which requires less memory and is efficient, and good results on standard benchmarks.
  • 2. What are the strengths of the paper? Clearly explain why these aspects of the paper are valuable.
    • - The paper is clearly written. The approach is simple, intuitive, and seems easily reproducible. Different design decisions are well motivated.
      - The method performs well on standard benchmarks.
      - Visualizations show the effectiveness of the approach.
  • 3. What are the weaknesses of the paper? Clearly explain why these aspects of the paper are weak.
    • The key contributions are the geometric net architecture and CMPNN, however, there's no ablation analysis to understand either of them. There's not much intuition behind the network architecture itself. Can the authors provide more empirical insights? Example:
      - What if fewer or more layers are used in Geometric Feature Net; i.e., what if it's shallower or deeper? Similarly, what if it's wider?
      - What is the contribution of Residual blocks?
      - What's the impact of changing the final feature dimensionality? How well does Nx32 perform or Nx024 perform? What's the speed/accuracy trade-off?
      - How important is CMPNN? For a synthetic setup, how does MPNN and CMPNN compare in terms of speed, memory, and accuracy?
      - What is the overall memory footprint and inference time? How does that compare against baselines? What's the impact of above experiments (e.g., deeper/shallower, wider) on memory and inference time?

      Generalizability:
      - Can the authors comment on the assumption on the problem setup? For example, current methods have N points in both images. But the approach seems generalizable to N and M points in different images and to one-to-many mappings.
      - Can the method be applied to the problem setup of [7]?
      - How does the method perform on dataset from "Dense Semantic Correspondence where Every Pixel is a Classifier" (and against the method)?

      Minor typos:
      L245: A -> a
      L843: 7 -> Fig. 7
  • 4. Paper rating (pre-rebuttal)
    • Borderline
  • 5. Justification of rating. What are the most important factors for your overall recommendation?
    • I like the simple idea presented in the paper. However, I believe that more ablation analysis is needed to tease out the impact of stated contributions.
  • 10. Final recommendation based on ALL the reviews, rebuttal, and discussion (post-rebuttal)
    • Weak accept
  • 11. Final justification (post-rebuttal)
    • After going through other reviews and the rebuttal, I upgrade my rating.
Reviewer #2

Questions

  • 1. Summary. In 3-5 sentences, describe the key ideas, experiments, and their significance.
    • In this paper, the authors proposed a a graph neural network which can transform weak local geometric features into rich local features for addressing the NP-hard assignment problems. In the experimental section, the authors make some experiments of feature matching both on synthetic dataset and real-world dataset with the proposed model and get good results. It is the most significant contribution that the traditional NP-hard assignment problems are replaced with a simple linear assignment problem with their CMPNN structure.
  • 2. What are the strengths of the paper? Clearly explain why these aspects of the paper are valuable.
    • Authors proposed a novel Compositional Message Passing Neural Networks to get an accelerated version of MPNN. Instead of associating the type for each edge manually, the authors use a neural network to predict the type of the edges and meanwhile explain their network Architecture in such detail. Moreover, this paper presents the comprehensive and enough experiments to varify the proposed framework.
  • 3. What are the weaknesses of the paper? Clearly explain why these aspects of the paper are weak.
    • Although the experiment for this CMPNN structure is enough, it still lacks some contrast experiments. I don't think there is no existing work on the matching problem that uses only the coordinates of feature points. Meanwhile, there exists some absent explainations for the key concepts.In additional, I suggest authors to enhance the qualitative discussion.
  • 4. Paper rating (pre-rebuttal)
    • Borderline
  • 5. Justification of rating. What are the most important factors for your overall recommendation?
    • The novelty and effectiveness of the proposed method and whether the author could explain their method clearly or not.
  • 6. Additional comments.

    • (1)Authors said:"For deep learning based methods, as far as the authors know, there is no existing work on the matching problem that uses only the coordinates of feature points."(line 524-526)
      You shouldn't be so sure about that. In additional, you'd better make some contrast experiments with other similar methods.
      (2)In figure 5 of Accuracy-Separation, why does the accuracy of your method remain stable with the increasing of separation?
      (3)In Table 1, could you give some explainations that why the matching accuracy of boat, bus, dining-table, potted-plant,train and tv-monitor are lower than other methods?
      (4)In figure 5 of Time-Separation, why does the curve sudden growth at the separation of about 60?
  • 10. Final recommendation based on ALL the reviews, rebuttal, and discussion (post-rebuttal)
    • Weak accept
  • 11. Final justification (post-rebuttal)
    • Parts of key issues have been handled.
Reviewer #3

Questions

  • 1. Summary. In 3-5 sentences, describe the key ideas, experiments, and their significance.
    • This paper presents a deep learning method for finding correspondences among sets of 2D coordinates in two images given only the coordinates (and not appearance-based features). The main idea is to use a graph neural network to take a set of coordinates of feature points from an image as input and output local geometric features that can be efficiently matched to the local geometric features of a second image using the Hungarian/Munkres algorithm. Tests on a synthetic dataset, the CMU House dataset and the PF-Pascal dataset show that the algorithm is reasonably fast and compares favorably with competing algorithms especially in terms of robustness to rotated feature points.
  • 2. What are the strengths of the paper? Clearly explain why these aspects of the paper are valuable.
    • The main strength of the paper is the novel idea of using a deep network to learn a set of local geometric features that capture the geometric structure of a local neighborhood of feature points and allow for efficient matching. The resulting features allow the matching algorithm to be efficient (Hungarian algorithm) and avoid the exponential inference problem that many competing algorithms face.

      The experiments show that the algorithm is faster than competing algorithms such as Factorized Graph Matching and Branch and Bound and is more robust to noise and outliers as well as rotations of the feature points for most object classes.
  • 3. What are the weaknesses of the paper? Clearly explain why these aspects of the paper are weak.
    • The explanation of compositional message passing networks (CMPNN) could be clearer. For example, the concept of a "type" for each edge is not well explained. Is the type essentially just a way of ordering the edges?

      In the results on real image matching using the PF-Pascal dataset, the lack of further discussion or analysis of the results is a weakness. The results (Table 1) show that under in-plane rotations of the feature points, the matching accuracy for some feature classes gets much worse while most suffer a small/modest degradation in accuracy. Some explanation for this would be helpful. Why are some object classes affected so much more than others?
  • 4. Paper rating (pre-rebuttal)
    • Weak accept
  • 5. Justification of rating. What are the most important factors for your overall recommendation?
    • The approach is novel (to my knowledge), practical and leads to very nice results, however the explanation of CMPNN is not very clear. Addressing the weaknesses mentioned above in the rebuttal would be helpful.
  • 10. Final recommendation based on ALL the reviews, rebuttal, and discussion (post-rebuttal)
    • Weak accept
  • 11. Final justification (post-rebuttal)
    • After reading all of the reviews and the rebuttal, I think the authors did a good job of addressing the reviewers' concerns. I think the paper is interesting, has novelty, and shows good results. I'm still in favor of acceptance.
Go Back
© 2019 Microsoft Corporation About CMT | Terms of Use | Privacy & Cookies | Request Free CMT Site
CMT Support