Superpixel Image Classification with Graph Convolutional Neural Networks Based on Learnable Positional Embedding

Round 1

Reviewer 1 Report

The content of the article corresponds to the topic of the journal. The article is relevant, dedicated to a new approach to the problem of classification of superpixel images based on GCNN.

The article clearly formulates the problem of classifying superpixel images, which is solved using a Graph Convolutional Neural Network. The paper uses three convolutional graph models (the Chebyshev graph convolutional network, graph convolutional network and graph attention network) to test the performance of the proposed model on different datasets of reference images.

The novelty of the paper is that the proposed IMGCN-LPE method makes GCNNs more powerful by expanding various graph structures by continuously learning the positional property and improving performance using ArcFace and cross-entropy as loss functions.

References to publications are related to the nature of the research conducted. Their novelty and quantity meet the requirements.

The presentation of the material is logical, accessible for perception, based on the axiomatics of Graph Convolutional Neural Networks. The drawings presented in the article are appropriate and visually reveal the investigated problem.

Author Response

This paper shows the Leanable Positional Embedding mechanism not only be applied to original graph structural data but also to graph represented data. Thus, our proposed method can help Graph Convolution Neural networks understanding the various arbitrary graph structures so that improve and get a good results in computer vision task. Anyway, thank you for your review of our paper positively and for taking your precious time.

Reviewer 2 Report

The subject is important, and it was well worked out.

Nonetheless:

1. What is written in lines 15 to 17 seems to me to contradict what is written in lines 31 to 37.

2. The content between lines 31 to 43 must be accompanied by supporting references.

3. There must be some introductory explanation between items 2 and 2.2.

4. The same between items 4 and 4.1.

5. Item 4.3 is very poor in arguments.

6. Conclusions must have at least some numerical values to support them.

Author Response

1. What is written in lines 15 to 17 seems to me to contradict what is written in lines 31 to 37.

We explained graph convolutional neural networks based on CNN convolution is not work well in computer vision task in lines 15 to 17, but normally CNN is powerful in lines 31 to 37. It is really weird that there is not enough reason to back it up. So we add the explanation why convolution is difficult to be applied on graph and indicate the solution of this issue by designing the covolution operator of graph in lines 41 to 54.

2. the content between lines 31 to 43 must be accompanied by supporting references.

There are not enough references to support our information so that we think it causes lack of reliability. So we add the references related to the explanation.

3. There must be some introductory explanation between items 2 and 2.2.

We think section 2.2 is not clearly to serve as the motivation of the proposed method. So, in lines 105 to 111, we changed the name of subsection(2.2) and presented the reason why this section is needed to explain and briefly said it.

4. The same between items 4 and 4.1

It needs the summarized introduction about section 4. So we present the briefly information of each sub-sections.

5. Item 4.3 is very poor in arguments.

we don`t recognize but thank you for checking it. We revised the initial letter to meaningful word (L means the number of stacking Layer, H means number of heads for attention mechanism) in whole tables.

6. Conclusions must have at least some numerical values to support them.

we add the numerical value of our experiment results in lines 409 to 413 for supporting and understanding our conclusion easily.

Reviewer 3 Report

This paper proposed a graph convolutional network with learnable positional embedding applied on images (IMGCN-LPE) that initialize positional information through a random walk algorithm and continuously learn the additional position-embedded information of various graph structures represented over the superpixel images.

The strong point of the paper:

1. This paper applies IMGCN-LPE to three graph convolutional models (the Chebyshev graph convolutional network, graph convolutional network, and graph attention network) to validate performance on various benchmark image datasets to validate the effectiveness of the proposed method. 

Weak points of the paper:

1. This paper is more like a technical report than a research paper and not enough novelty. It looks like a combination of different techniques and list the experimental results without much analysis. 

2. The relative work part talks a little too much of the details, i.e., the mathematical formula, but does not clearly stress the current issue of the area to serve as the motivation of the proposed method.

3. The presentation is not of good quality. For example, figure 3 is distorted

and a little blurred, the terms like GCN-1L, and GAT-1H in experimental results are not clearly explained. 

 

 

Author Response

1. This paper is more like a technical report than a research paper and not enough novelty. It looks like a combination of different techniques and list the experimental results without much analysis.

It is really good point and we agreed that applied various expressive idea and techniques to get a good results. But, what we wanted to say in this paper is finding the reasons about graph convolutional neural networks (GCNNs) limitation of computer vision task and especially be improved the performance on superpixel-based image classification. Generally, idea and techniques that we used were applied on original graph structural data (social networks, chemical or biological molecules etc..) and worked really well. But there were not good result and not enough studies of computer vision task because of inevitable issues which cannot be solved on GCNNs in general. So, We tried to solve them by combining various ideas and get a results better than previous works. This is the biggest contribution of this paper, we think.

2. The relative work part talks a little too much of the details, i.e., the mathematical formula, but does not clearly stress the current issue of the area to serve as the motivation of the proposed method.

we explained the graph convolution idea according to the mathematic formulas but it makes the information more complicated. So, we explained the introduction briefly in lines 107 to 110 so that emphasized the graph convolution is the most important concept of this paper. Also, we used basline models of two types of graph convolution : spectral-based Chebnet, and spatial-based GCN, GAT. So, we explained the basic information about them in section 2.2.

3. The presentation is not of good quality. For example, figure 3 is distorted and a little blurred, the terms like GCN-1L, and GAT-1H in experimental results are not clearly explained

we didn`t know but thank you for checking them. Thanks to your review, we revised the figure2, 3 more clearly and deleted the blur or distorted part and make the arguments (L, H) change and explained clearly in whole tables.

4. English language and style are fine/minor spell check required

We re-checked the word spells again.