Questions about YOLOV3 - Githubissues

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.75k stars 7.96k forks source link

Questions about YOLOV3 #5513

Open ardeal opened 4 years ago

ardeal commented 4 years ago

Hi,

I got 2 questions about YOLOV3:

Question 1: on the paper of YOLOV3, Table 1 Darknet-53, the number of convolution layers is 52, but not 53. is Connected 1000 a convolution layer? or any other explanation?
Question 2: I checked the network structure in file yolov3.cfg(https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov3.cfg). The network before line 549 is the same as that on YOLO V3 paper. But the network after line 549 is not explained on YOLOV3 paper. Could you please point out where the network after line 549 is explained?

Thanks, Best Regards,

AlexeyAB commented 4 years ago

Table-1 in the Yolov3 paper is wrong, since Darknet-53 uses [convolutional] layer instead of [connected] layer at the end of model: https://github.com/AlexeyAB/darknet/blob/master/cfg/darknet53.cfg#L557-L562
Yolov3 used FPN https://arxiv.org/abs/1612.03144

ardeal commented 4 years ago

Hi AlexeyAB,

Many thanks for you reply!

2 more questions:

Yolov3 used FPN. where is this explained? is it explained on Yolov3 paper? I would like to know how it is used. I would like to read the detailed explanation on paper.
According to my 2 questsions(Darknet-53 is wrong, FPN is not explained on Yolov3 paper), it is impossible to reproduce Yolov3 performance according Yolov3 paper merely, right?

Thanks & Best Regards, Ardeal

ardeal commented 4 years ago

@AlexeyAB Hi AlexeyAB,

Do you have any update about those questions?

Yolov3 used FPN. where is this explained? is it explained on Yolov3 paper? I would like to know how it is used. I would like to read the detailed explanation on paper.

According to my 2 questsions(Darknet-53 is wrong, FPN is not explained on Yolov3 paper), it is impossible to reproduce Yolov3 performance according Yolov3 paper merely, right?

Thanks and Best Regards, Ardeal

ardeal commented 4 years ago

@AlexeyAB, Hi, Do you have any update?

AlexeyAB commented 4 years ago

YOLOv3 predicts boxes at 3 different scales. Our system extracts features from those scales using a similar concept to feature pyramid networks [8].

https://arxiv.org/pdf/1804.02767.pdf

ardeal commented 4 years ago

@AlexeyAB Hi,

This section just mentioned it used FPN, but it didn't explain how and where FPN would be used in Yolov3. If I read this section, I cannot know how FPN is used/connected in YOLOv3. I cannot write code to implement that.

Is there any more detailed description about how and where FPN is used in YOLOv3?

Thanks, Ardeal