search

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

YOLO Algorithm #1330

Open skipm3 opened 6 years ago

skipm3 commented 6 years ago

Hi there. This is not an issue btw i just want to ask, what is algorithm used by YOLO? Any reference? Thank you.

AlexeyAB commented 6 years ago

https://pjreddie.com/media/files/papers/YOLOv3.pdf

skipm3 commented 6 years ago

@AlexeyAB is there any name for algorithm what YOLO used? Like haar cascade or something?

AlexeyAB commented 6 years ago

This is single-shot Convolutional Neural Network that is part of Deep Learning.

kooscode commented 6 years ago

@skipm3 - What AlexeyAB said and you can also look at https://leonardoaraujosantos.gitbooks.io/artificial-inteligence/content/single-shot-detectors.html

skipm3 commented 6 years ago

@AlexeyAB @kooscode hi. i've read some things about CNN here in layer section.

  1. Does this repo use same method?
  2. If it is, what width 32 and height 32 means? Does the input image is resized to 32x32?
AlexeyAB commented 6 years ago

@skipm3

  1. yes
  2. yes, any image will be automatically resized. For CIFAR-model to 32x32. For Yolo-model to 416x416.
kooscode commented 6 years ago

@skipm3 - thats correct, it will resize the image equal to the network input size.

you can resize the network input size for YOLO easy by just changing the .cfg file for network definition.

but beware that larger network input size has exponential effect on memory and gpu utilization for training and inferencing.

BUT it helps a lot with detecting small items on larger images. i.e. we use YOLOV2 with a network input size of 512x512 for detection of small artifacts in aerial drone imagery.

but training requires lots of GPU memory and increased training time..

skipm3 commented 6 years ago

@AlexeyAB @kooscode I got this image image Does YOLO just runs convolutional network? I mean, look at this. This is a ConvNet architecture: image and this image So, YOLO just input -> conv -> result?

AlexeyAB commented 6 years ago

@skipm3 Yes, as I said: https://github.com/AlexeyAB/darknet/issues/1330#issuecomment-410213365

This is single-shot Convolutional Neural Network that is part of Deep Learning.