yolo in caffe

[Jeremy Rutman]

Has anyone gone and implemented yolo in caffe?

I ran across  a fork https://github.com/handong1587/caffe-yolo  but there are no instructions or docs it seems.

Is it an active branch? Last update was Dec. 2015.

[Dmytro Prylipko]

I am also interested in YOLO for Caffe. Is it usable?

[Yassine Bezza]

Hello,

I found this but I did not test it yet.

https://github.com/MelodyHai/caffe-yolo

[Xing Wang]

You may want to check this out 

在 2016年3月21日星期一 UTC-7上午2:42:34，Jeremy Rutman写道：

[Hugo Feng]

Hey Wang,

Impressed! In terms of the weight file you mentioned, how do you convert from the darknet model? Anything worth a heads up?

Since darknet inputs distributes from -1 to 1, while caffe uses mean image substraction, so how do they fit into the convertion? Any preprocessing you added?

Thank you!

Hugo

在 2016年5月21日星期六 UTC+8上午8:34:47，Xing Wang写道：

[Xing Wang]

Hi, Hugo,

Thanks ! 

If you use caffe io to load image, in fact, the loaded image is in [0, 1], quoted from caffe io 

"While Python represents images in [0, 1], certain Caffe models like CaffeNet and AlexNet represent images in [0, 255]"

You can find more details here https://github.com/BVLC/caffe/blob/master/python/caffe/io.py

Mean image subtraction is  necessary for ImageNet models, but not here. 

According to the yolo source code, I think yolo only requires the input image to be in [0, 1], but not [-1, 1].

For caffemodel part, I directly read the darknet model and stored it into a caffemodel without any modification.

Xing

在 2016年5月24日星期二 UTC-7上午1:39:44，Hugo Feng写道：

[yaxi Wu]

Hi, Xing,

Thanks for your yolo in caffe. It is really wonderful!

I haven't figured out how to convert darknet model into a caffemodel. Did you use a tool or write a script to complete this? Would you give me some advice about it?

Thank you!

Yaxi 

在 2016年5月25日星期三 UTC+8下午1:24:17，Xing Wang写道：

[Hugo Feng]

Thank you Wang! That helps me a lot~

在 2016年5月25日星期三 UTC+8下午1:24:17，Xing Wang写道：

Hi, Hugo,

[Xing Wang]

Hi, Yaxi,

I updated the repository and added a create_yolo_caffemodel.py to create caffemodel files from yolo's (.weights) files.

Hope it helps.

Xing 

在 2016年5月26日星期四 UTC-7上午8:28:20，yaxi Wu写道：

[yaxi Wu]

Hi, Xing,

Thank U!  It really helps me a lot.

Yaxi

[ofir liba]

Hi Xing,

I'm trying to convert AlexNet from Darknet to Caffe by using your script "create_yolo_caffe_model.py" . 

To do that I need to convert first alexnet.cfg to caffe prototext file. could you explain how you did the conversion for Yolo ( cfg-->prototext) ? when I do it , i encounter some problems.

 

Tx,

Ofir

[banerje...@gmail.com]

Hi Xing,

Thank you so much for the scripts -- it is excellent. Can you please explain the line:

offset = np.transpose(np.reshape(np.array([np.arange(7)]*14),(2,7,7)),(1,2,0))

in yolo_main.py

I am confused with the number 14.

Thanks and regards,

Serene

[Lucas Thom]

Hey Xing,

Is this right?

probs = np.zeros((grid_size,grid_size,num_box,num_class))
class_probs = np.reshape(output[0:(grid_size*grid_size*(num_box*5+num_class)/3*2)],(grid_size,grid_size,num_class))
scales = np.reshape(output[(grid_size*grid_size*(num_box*5+num_class)/3*2):(grid_size*grid_size*(num_box*5+num_class)/3*2)+grid_size*grid_size*num_box],(grid_size,grid_size,num_box))
boxes = np.reshape(output[(grid_size*grid_size*(num_box*5+num_class)/3*2)+grid_size*grid_size*num_box:],(grid_size,grid_size,num_box,num_box*num_box))
offset = np.transpose(np.reshape(np.array([np.arange(grid_size)]*grid_size*num_box),(num_box,grid_size,grid_size)),(1,2,0))

[NM]

Hi,

Does this now have training included as well? Im looking to train YOLO on caffe 

Thanks 

[Rishabh Gupta]

Hey all so I followed steps from https://github.com/xingwangsfu/caffe-yolo using cfg/yolov1/yolo.cfg  but I encounter the following error 

ERROR:root:local layer is not supported
ERROR:root:detection layer is not supported
can anyone please help me with that .
Thanks

[tone...@gmail.com]

Someone may have a look at https://github.com/yeahkun/caffe-yolo and https://github.com/choasUp/caffe-yolo9000 , if you prefer not to use Python.
However, their projects are flawed. At least they have achieved their goal by modifying Caffe itself, and this will be a problem with many users by (1) implicitly fixing the version YOLO can apply to to 1.0.0-rc3 instead of the newest version possible; (2) using many POSIX-onlly functions that may fail to compile under Windows.

[Lin Jay]

i found someone implemented yolov2 on caffe version, seems which is reference from yeahkuns work on v1.

from https://github.com/karta0807913/caffe-yolo-9000 it works fine for me.


Jeremy Rutman於 2016年3月21日星期一 UTC+8下午5時42分34秒寫道：

[zenge...@gmail.com]

@Lin Jay. hi

Can this project train the yolov2 network? and what about the accuracy compare to darknet's training?

在 2017年7月15日星期六 UTC+8上午12:32:52，Lin Jay写道：