An Efficient Multi-Scale Anchor Box Approach to Detect Partial Faces from a Video Sequence
Round 1
Reviewer 1 Report
This paper proposed the method of correctly detecting partial faces from videos. The proposed method is tested on the popular and challenging face detection benchmark, namely Face Detection Data Set and Benchmark (FDDB) dataset. However, it is difficult to find technical novelty in current paper. Detailed descriptions of the proposed methods and techniques are required. What techniques and strategies have been specifically developed to detect partial faces? How to build a deep network structure for accurate and fast partial face detection?
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The topic addressed by the authors is interesting, but some flaws can be found in the presentation of the work affecting the comprehension of the study. The dissertation seems confusing at times, particularly:
- English editing would be necessary.
- Introduction and scope. It is not clear if the authors intended to treat real-time applications or not, since they generically talk about “photograph”. This introduction should be more consistent with the declared scope of the paper (i.e., Contributions, Novelty).
- Background study. The logical thread of the paragraphs is unclear, they seem more like notes than parts of a paper. It would be better to avoid starting paragraphs with a bulleted list (line 117)
- Methods and experiments. The authors state: “The number of classes is changed considering the FDDB dataset, and only one class label is fed, i.e. the face.” Which is the logical thread with the previous paragraphs? What do the authors refer to by “the number of classes is changed”? Classes of what? One can guess that authors are talking about a convolutional neural network, but which one? Moreover, acronyms should be written in full the first time they appear in the text.
The scope of the work is stated at the beginning of the article, but the following dissertation is unclear; as said before, there seems to be no logical thread between the paragraphs.
- Proposed work. Authors talk about figures 4 and 5, but which was the context of the comparison? Were the different methods compared on occluded faces? Are authors already working on real time?
- Overall accuracy on testing videos. Could the authors insert images from experiment?
At line 284 authors talk about “things”, but weren’t they talking about faces?
- A discussion paragraph is missing, so conclusion should be enriched and expanded, also deepening applications of the proposed method and possible future outcomes, in addition to a clearer dissertation of the results.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Reviewer 3 Report
- The authors should include the detail about the proposed work. such as system flow diagram, algorithm, etc...
- Should also clearly compare the performance (accuracy) of the proposed work with the other existing methods.
- The authors described that " The existing models fail to detect
the partially covered faces" in the abstract. Therefore, the authors should create the dataset that contains partially covered faces for evaluating and comparing the performance of proposed work with the other existing methods. - Overall, the authors need to improve the presentation design about proposed work and need to do more experiments to confirm the effectiveness of the proposed work .
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The authors sufficiently improved the quality of the submitted work and of the presentation itself.
Reviewer 3 Report
FDDB is a popular benchmark dataset for face detection and a lot of research works has been done using that dataset as described in this URL: http://vis-www.cs.umass.edu/fddb/results.html
Therefore, the authors should compare and analyze their results with some of the state-of-the-art methods by using the same performance evaluation method such as discontinuous and continuous ROC.
