BFE-Net: Bidirectional Multi-Scale Feature Enhancement for Small Object Detection
Round 1
Reviewer 1 Report
This paper proposes the Bidirectional Multi-scale Feature Enhancement Network (BFE-Net) based on RetinaNet. After introducing a bidirectional feature pyramid structure to shorten the propagation path of high-resolution features, the paper utilizes residually connected dilated convolutional blocks to fully extract high-resolution features of low-feature layers. Then the paper supplements the high-resolution features lost in the high-level feature propagation process by leveraging the high-level guided lower-level features. The authors showed some experimental results to improve some previous state of arts. The paper is well organized. But I have a few questions.
- Equation 3. The term F dilated (i) is appeared in both left and right sides of the equation.
It is not logical definition of an equation. Please check.
- What is the definition of Fconcat in equation (1)?
- Are D(C2 ) in equations (1) and (4) the same?
Apart from these, I have on other objections.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
===== Synopsis:
The study improves the RetinaNet network model to specifically detect
smaller objects better. The improvment is marginal. The paper is
clearly written.
===== General Comments:
I enjoyed reading the study because it has a very concise and clear
introduction and the objectives are also well formulated.
I find that it misses only one point of analysis and discussion: how
much slower is the improved network model, in comparison to the
original? (or perhaps I missed that aspect in the manuscript)
The improvement for small objects (2.8%) is larger than the general
improvement (2.1%). One could argue that the model was merely biased
toward smaller objects, as the targert proportion for smaller objects
is also larger.
===== Specific Comments:
line 152: atrous convolution. atrous?
line 153: global and local...sentence appears incomplete.
line 175: 'but its semantic information is less and is easy to be disturbed by noise
points'. does not sound right somehow.
line 179: FPN not introduced. you cite it later at line 183.
line 185: 'so that extract features more fully for better detection'. doesn't sound right.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
