Multi-Scale Kolmogorov-Arnold Network (KAN)-Based Linear Attention Network: Multi-Scale Feature Fusion with KAN and Deformable Convolution for Urban Scene Image Semantic Segmentation
, Shuo Liu 1, Jie Wu 1, Weichao Sun 1, Qingke Wen 1, Yibiao Wu 2, Xiujuan Qin 2 and Yanyou Qiao 1
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsPlease see the attachment.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper presents the MKLANet network, which features an innovative attention mechanism. Some detailed comments are as following:
(1) The description of the model could be more detailed, particularly regarding the specific mechanisms of the MKLA and LR CED modules. A deeper exploration of these components would strengthen the theoretical foundation of the work.
(2) The results for edge accuracy and small object segmentation across different datasets are not presented with targeted comparisons. It would be beneficial to include more visual analyses in the ablation studies to fully showcase the advantages of the various components of MKLANet.
(3) More comparative experiments with the latest and most advanced models could strengthen the comprehensive evaluation of MKLANet's performance. For instance, though MKLANet shows significant improvements over the recent Mamba-based algorithm RS3Mamba, it would be more convincing if including comparison with the latest Transformer-based methods.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIn this work, a segmentation network named MKLANet, based on Kolmogorov-Arnold Networks for attention is proposed. The main contributions are: 1) a deformable convolution block, and 2) the parallel implementation of a multi-scale KAN linear attention network to extract local details and a Long-range Cascade Encoder Decoder to extract global details.
Overall, the article is well structured. The evaluation of the proposed model on 4 different datasets is very attractive.
Some minor remarks:
- the title seems a bit long, I suggest that the authors reduce the size of the title (it is necessary to define what is used in linear attention? deformable convolution? or urban scene?)
- in line 81 there is a typing error
- line 124 typo? (both)
- in image 2 they mention ‘Offests’, isn't it ‘Offsets’?
- line 190, possible typo
- line 212, A-?, I think it's A^-. The same for B and C
- line 215, D is your contribution? if not, what parameter does it correspond to in the SSM equations?
- line 235 mentions that x_i is augmented in equation (11), are you sure this is the correct equation?
- a typo on line 307
- the intention of this citation is not clear: the convolution block is defined in the referenced works? or are they used in them?
- I suggest checking the quality of the formatting in all the equations.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsNo more comments
