ucas-vg / effective-fusion-factor Goto Github PK

Hi, I'm very inspired by your paper. It's very novel and logical! But I am confused by some sentences in the paper.

1. When you explain Eq. 3 in 3.4 section, you mention that,

"While applying bigger α₃⁴, C4 will acquire more information used for detection task of shallow layers, and lost more information used for detection task of deep layers, vice versa."

But in the following paragraph, you mention that giving up part of information for detection of deep layers is by applying small α₃⁴. I think these two statements are contradictory. And in other parts of the paper, I think statements about α₃⁴ are identical to the latter. I understand the latter, but I am very confused about the former.

2. It's mentioned in 3.3 section that Tab. 2 shows the failure of the initialization of f_inneri and f_layeri. But it is shown in Tab. 2 that the two initialization methods can both increase AP₅₀^tiny and decrease MR₅₀^tiny. Why are they failures?

3. In 3.1 section, it's mentioned that,

The stratification of CityPersons in FPN is still similar to that of Tiny CityPersons since the anchor of Tiny CityPersons is also reduced by four times.

I can't understand why the stratification of CityPersons is still similar to that of Tiny CityPersons since the size of anchors are reduced, focusing on the shallower layers.

These questions have confused me for several days. Could you please help me and explain them in a more detailed way? Thank you very much!

Hello author！In this paper, I am very inspired by your insights about the influence of different feature maps on small target detection, but I have a question about gradient equation in the paper.

In the gradient formula of C4, the backpropagation gradient information of P2 layer should pass through the fusion factor between P'2 and P'3. Why is it not reflected in the formula?

I am not very clear about this question or can you please provide appendix mentioned in this paper?

Hi~, authors, I have two questions about this code.
Firstly, is the code published now complete? I can not find some functions or python file.
Secondly, in "train_net.py", line 30, the "build_detection_model" function needs two arguments. One is cfg and another is fusion_factors. But there is only one cfg argument in the code. I have some doubts about this.
Looking forward to your early reply.