Darkflow: How to setup structure of own dataset

Created on 18 Jul 2017 · 4Comments · Source: thtrieu/darkflow

First of all great work! I have many traffics sign images which has been categorized in folders of name of their type.I need to know how can I setup my images for training means:

How can I make annotation is there any tool for it?and where should i place the images and annotation.
I have followed the steps in section Training on your own data set ,I have around 105 folders of different sign which mean my classes will be 105 and I have updated label.txt with 105 names.When I am making new cfg file from where i will get weights for it?

flow --model cfg/tiny-yolo-voc-3c.cfg --load bin/tiny-yolo-voc.weights --train --annotation train/Annotations --dataset train/Images

Or it will be done through tiny-yolo-voc.weights?

Source

ManojPabani

Most helpful comment

Hi @merryHunter
How about image size?
is there any limitation of image size?

I actually want to use my detection to detect things that is slightly different in color (the size is quite and the color is quite similar)
Can I still use YOLO for that? or does it mean I will need a lot of images because they are just different in color?
Need Advice

normansiboro on 23 Nov 2017

👍2

All 4 comments

Hi @ManojPabani

Try to use for example this one: https://github.com/tzutalin/labelImg
As if you followed the tutorial, you may have noticed that all images are placed in the one folder, as well as annotations to them.
If you want to use weights, you need to download them from yolo website. Otherwise, you can start training from scratch by using another command:
# Completely initialize yolo-new and train it with ADAM optimizer
flow --model cfg/yolo-new.cfg --train --trainer adam --annotation train/Annotations --dataset train/Images

merryHunter on 12 Aug 2017

Hi @merryHunter
in your command, do we still need to put label parameter (--label) ?
and
is there any consideration of how many pictures are enough to represent a class?

normansiboro on 22 Nov 2017

Hi @normansiboro
--label param only makes your life easier. It's better always specify labels path file explicitly, so you don't confuse them. I particular, for your own dataset you should change classes in the default labels.txt or in the one where you specify the path.
For pictures - a lot depends also on the network and input size, image quality as well. Generally, you can take a look how many samples per class Pascal VOC contains. I'd say you'll need at least 500 samples labeled in order to learn something.

merryHunter on 23 Nov 2017

👍2

Hi @merryHunter
How about image size?
is there any limitation of image size?

normansiboro on 23 Nov 2017

👍2

Was this page helpful?

0 / 5 - 0 ratings