YOLO V2 paper question - Githubissues

AlexeyAB / darknet

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

http://pjreddie.com/darknet/

Other

21.64k stars 7.95k forks source link

YOLO V2 paper question #883

Open 89douner opened 6 years ago

89douner commented 6 years ago

In "YOLO9000, Better, Faster, Stronger" paper, they mentioned that "the model resizes 448 to 413 (drawing final feature map 13x13) for recall improvement". Also, In the paper, they attached "Table 6: Darknet-19" showing 448 (drawing 7x7).

I don't understand why the final feature map size is different???

AlexeyAB commented 6 years ago

Darknet-19_448 isn't the same as Yolo v2.
In the table-6 is shown Darknet-19_224 (not Darknet-19_448) https://arxiv.org/pdf/1612.08242.pdf

Darknet-19_448 - input_size=448x448, output_size=14x14
Darknet-19_224 - input_size=224x224, output_size=7x7
Yolo v2_416 - input_size=416x416, output_size=13x13

89douner commented 6 years ago

@AlexeyAB Does the table-6 represents Darknet-19_224 (input size: 224x224)? the table shows first output 224x224, so I think ... Doesn't the table-6 represent Darknet-19_448 (input size:448x448) ??? maybe.. Does Darknet-19_224 resize input image 224x224 to 448x448 ??

AlexeyAB commented 6 years ago

Does the table-6 represents Darknet-19_224 (input size: 224x224)?

Yes.

https://github.com/AlexeyAB/darknet/blob/6390a5a2ab61a0bdf6f1a9a6b4a739c16b36e0d7/cfg/darknet19.cfg#L4-L5