Attentive Part-Based Alignment Network for Vehicle Re-Identification
Round 1
Reviewer 1 Report
The paper introduces an attentive part-based alignment network for vehicle re-id. The method is comprised of several components, which are verified in the experiments. The method sounds reasonable and the paper is well-written. However, there are some issues that should be addressed:
- It is not clear to me how the parts are generated in the networks, e.g., in Eq.2, how are the N_p part features determined?
- What is the feature split operation in Fig. 2?
- The loss function in Eq.14 consists of four terms. How are the coefficients determined? How will they affect the model performance?
- Neural attention is widely studied and the discussion in Sec 2.2 should be expanded to include some recent works, e.g., Cascaded parsing of human-object interaction recognition, Matnet: Motion-attentive transition network for zero-shot video object segmentation, and Group-Wise Learning for Weakly Supervised Semantic Segmentation.
- It will be better to provide some visual examples of the parts learned by the network.
- The network is very complex with many components, however, its performance does not outperform existing methods, as reported in Table 7. More analysis should be provided regarding the results.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
This paper proposes a Multi-receptive-field Attention (MA) module to improve feature representations. The paper is overall well written. In addition, the authors conduct extensive experiments on two popular vehicle Re-ID datasets. I have some review comments to improve the quality and readability of the paper.
1) There are unnecessary descriptions in the abstract. The authors need to point out the main advantages of the proposed work clearly in the abstract.
2) The authors show that APANet outperforms all other GM-based methods. To clearly address this achievement, they need to describe the reasons in detail.
3) The authors need to clearly describe their training settings and test settings to improve the readability.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Thank you for submitting the paper. However, some improvment actions should be taken:
- The authors have to present research methods under the abstract;
- The number of keywords should be extended;
- The explanation shoud be given for VeRi-776, VERi-Wild, Res2Net and other abbrevations. Abbrevations have to be explained befiore they are mentione in text for the first time.
- By the end of introduction the authors could present the structure of their paper.
- At the beginning of section 3, please, place more valuable information. It is recommended to present here research methodology, research questions and methods.
- The names of sections 3.1-3.4 should be unified. I suggest to use the style present at 3.2-3.3 for naming 3.1 and 3.4. Such as 3.1 Attentive Part-based alignment (APANet) and network structure; 3.2 Part-level Orthogonality Loss (POL); 3.3 Part-level Attention Alignment Loss (PAAL); 3.4 Multi-receptive-field Attention (MA).
- Not all dimensions in formulas are explained, such as Lpo in equation 2, Dc in equation 3, Cout in equation 4, Lpaa, K, P in equation 5, Cin, r in equation 8, Ls, j in equation 10, LTriHard in equation 11. Also, H and W in line 256.
- The authors have to add further research directions in the discussion section.
- The authors have to extend the conclussions section and present here their research limitations.
- The abbrevations in Table 7 should be explained below the Table.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The revision has addressed my concerns. I am happy to accept it.
Reviewer 3 Report
No comments