The First Step : Get the Data!

The very first thing to do is to take a lot of pictures of the data that you wish to train. Due to memory limitations, I had a relatively small dataset of around 70 images. 

Second Step : Labelling the Data

For this step, Label Studio needs to be installed on the machine. Installation instruction can be found on the Installation guide page.

The above video only shows one image being processed. You can create your dataset in this fashion by importing multiple images at a time and exporting them in a file. Each Image needs to be manually processed.


1. Make sure the object does not go out of the box even slightly. 

2. All the objects to be detected need to be marked in all the images, else the software might get confused. 

What are the Labels exactly?

The Label Studio generates a zip folder of 

1. Folder with all the images 

These are all the images that were uploaded

2. Folder with text annotations 

There are as many text files as the images that were uploaded. Each image has an associates text file with it, which has the same name as the image. It is necessary for this to be the case. The generated text file contains five numbers. The first number is the object class ID. The second is the object bounding box center in x axis, followed by object bounding box center in y axis, width and height. 

3. Classes.txt file

The classes.txt file contains the names of all the classes. Note that this file is not the same as the .names file. It needs to be converted in the .names format.