A Precision Efficient Method for Collapsed Building Detection in Post-Earthquake UAV Images Based on the Improved NMS Algorithm and Faster R-CNN

Round 1

Reviewer 1 Report

There are many vague and serious points regarding the submitted manuscript. First of all, as we know UAV data have high resolution and are beneficial in many ways for building damage detection recently. But if you go through the proposed method, I can not see any of those potentials are tackled! 

Secondly, according to EMS98, there are several damage classes and how it is discussed in your research work? How you can make a adopt those classes with the one you defined as "Collapsed Building"? How you treat other classes and how it is defined in you data set, both in your GT and results?

You mention in the abstract that the results are improved by 1 percent! Is this a good achievement?

There are many scientific issues regarding the submitted manuscript and it would be very good for the authors to study some well-know papers in building damage detection aside from the CNN based methods just to understand the damage detection problem in the first place.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Please add the equation (12) described from 359 to 364, which seems to be missing.

This interesting paper analyses high resolution UAV images of collapsed buildings from three different post-earthquake areas in China, improving one of the most popular deep-learning based object detection models for feature extraction, Faster R-CNN, by introducing Deformable Convolution (DCNv2) to extend its adaptability to the irregular geometric characteristics of collapsed buildings. The experimental results show that the improved model effectively addresses the issues of false detection and missing detection. Furthermore, instead of the traditional IoU, the authors propose the IPO (Intersected Proportion of Objects), thereby improving the NMS algorithm and evaluation criteria (Precision and Recall), validating the IPO strategy in the elimination of detected overlapping bounding boxes, and evaluating the detection of collapsed buildings.

We look forward to seeing small yet overlapping boundary boxes n the future, instead of large bounding boxes, that can improve the detection of collapsed buildings.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In this paper, the author proposed a collapse building detection. I am afraid this paper can not be published right now.

1. The novelty of the proposed method is not very clear.
2.  Is the only bounding box enough for practical application? Why not work out a segmentation of the boundary of collapsed buildings?
3. Can the data and codes be shared if the paper is published? 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I am satisfied with the answers

Reviewer 3 Report

In this paper, the author made a revision on the paper but did not resolve all my concerns. 

1. The novelty of the proposed method is quite limited.
2. The task in this paper is less challenging than the segmentation of the boundary. I don't think it is practical in the real application.