Feature Matching for Remote-Sensing Image Registration via Neighborhood Topological and Affine Consistency
Round 1
Reviewer 1 Report
This paper presents an efficient method to eliminate false matches.The topological relation of local matches and affine transformation consistency are regarded as important indexes to judge whether a pair of matches is correct.The existing methods also often use geometric distribution consistency of matches in local regions to eliminate false matches.The difference of the method in this paper is that the local topology constraint and affine transform consistency constraint are designed as penalty terms and added into the loss function. From the perspective of innovation, the paper is not significantly innovative.The experimental results show that this method can eliminate a large number of false matches efficiently and reliably. However, the authors need to explain the following two problems:
(1) In the experiments, satellite and aerial images (including UAV images) were used. In practical applications, the size of remote sensing image is very large. An UAV image usually has tens of millions of pixels, and a satellite image may have hundreds of millions of pixels. However, the images in the experiments are all small image blocks with only a few hundred thousand pixels, why not use the original large size images for validation? Will full-size images cover more complex scenes, such as areas with more repetitive structures, affect the performance of the proposed method?
(2) In the experiments, is the correctness of each pair of matches judged manually or automatically by designing an algorithm? Please provide necessary instructions in the paper.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The main comparative study in this work is LPM. However, descriptions about the LPM is not much. I suggest the authorsbriefly introduce LPM to give readers a clear understanding.
Although the authors provide visual comparisons between LPM and TAT in Fig.4 and 5, it is not clear why TAT can eliminate mismatch points that the LPM cannot eliminate. The authors could put some explanations in this paragrph.
As described in line 181, eq. (4) and (5) shows how the LPM quantizes the distance metric d. It seems like the authors also use the same rule as LPM, but the description here will make readers mistakenly think that eq.(3) is also part of LPM.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper presents a method for removing mismatches for robust feature matching on remote sensing images. The main idea is that between two images in the same or similar scene, the neighborhood topology of corresponding feature points should be consistent, and there are affine constraints in the local area of the image.The topic is interesting and matches well for Remote Sensing journal. The paper contains some review of related works. However the paper has some unclear points and the following minor concerns.
The authors describe the major contribution of the paper as follows:ï‚·
«1) We introduce local affine constraints into the neighborhood topology consistency approach. This method can deal with the problems of severe geometric distortion and obvious viewpoint change of images which can not be overcome by previous methods. 2) We formulate the problem as a mathematical model and derive its solution, and then conducted qualitative, quantitative and robust experiments on four remote sensing image datasets. Experimental results show that our method has better matching performance than other mainstream methods and can deal with matching problems in milliseconds.»
This «better matching performance» is confirmed mainly by Figure 8. However, in Figure 8, a significant (about 5% for recall and precision) advantage of the proposed method over the LPM method is observed only for one dataset (SAR) out of four considered. At the same time, the speed of the proposed method is an order of magnitude lower than the speed of the LPM method. Perhaps it makes sense for the authors to describe specific use cases for which their method can be applied. For this purpose, for example, section 4.3 (Qualitative Results) can be supplemented. The description of such specific use cases for their method can be based on the description that the authors have given for the SAR dataset:
«Because there is a lot of noise during the SAR imaging process, the quality of the image is not high and there is affine transformation, which brings great difficulties to feature matching. This type of image matching task usually occurs in positioning and navigation problems, that is, real-time matching of UAV images and satellite images stored in the database to obtain the current precise position.»
It would also be great if the authors explained how their theoretical innovation leads to such practical results.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors have answered all the questions. I recommend accepting the paper for publication.
Reviewer 2 Report
I have no further questions.
