An Intelligent Weighted Object Detector for Feature Extraction to Enrich Global Image Information
Round 1
Reviewer 1 Report
1. Please fix the typos, for example, line 29.
2. The authors have mentioned that they have developed a cheaper convolution method (line 56). Please explain which metric has been used to calculate that? Please also provide the cost analysis.
Author Response
We appreciate your questions and would like to respond to your comments and suggestions as follows:
Point 1: Please fix the typos, for example, line 29.
Response 1: Thank you for your careful review, we have corrected the errors in the paper.
Point 2: The authors have mentioned that they have developed a cheaper convolution method (line 56). Please explain which metric has been used to calculate that? Please also provide the cost analysis.
Response 2: The number of parameters of the model can be used as the evaluation standard. The following is an analysis of the specific convolution calculation as an example: in the convolution layer of YOLOX, when the size of the input is 128 × 80 × 80, the desired output size of 256 × 40 × 40, YOLOX practice is to use the size of 256 × 3 × 3, the step size of 2 convolution kernel for feature extraction of the input, which is also the general practice of other models, and the number of convolution kernel parameters at this time is 128×256×3×3=294912, while the method of cheap convolution is first to use the convolution kernel with the size of 128×3×3 and step size of 2 to extract the features and get the feature map x1, and then use the 128×1×1 convolution kernels with the step of 1 to extract the feature map x1 to get the feature map x2, finally, x1 and x2 are contacted according to the channel dimension to obtain the desired output, and the computation is (128×128×3×3) + (128×128×1×1) = 163,840, which shows that the parameters are nearly reduced by half.
Reviewer 2 Report
The article deals with the problem of increasing sensitivity and same time offloading computational burden to the CNN by introducing ghost typed features yielded by another convolutional feature derivate extracted from set of previously calculated by the mother NN. The results of implementation gives does not imply significance in quantitative manner as in qualitative. An introduction of another level of feature extraction, where an increase of feature sets information content utilizes not much local but from global descriptors often ortogonally set to the inner layer derivates of the modded NN, reflects that qualitative measure worth further research.
Despite of significance and interesting reading there is only one aspect that bothers me much and can increase quality of the work, so it must be improved. It is related to the thorough proofreading!
A thorough proofreading is mandatory! Although there are no many grammatical or typo errors, here-and-there minor misspells or editing mistakes occur.
Other aspects regarding scientific soundness, methodology, and presentation of the results are fine.
Author Response
Thank you for your comments. We will check the spelling and grammar of the article.
