Object Detection - Characters

Question

I am new to object detection, I have Tamil Language Characters and, I successfully done the classification part using CNN.

For object detection, I am following this example - https://towardsdatascience.com/getting-started-with-bounding-box-regression-in-tensorflow-743e22d0ccb3

We have to create the bounding box and create the annotation file which has the number of objects present in the image and the (Xmin, Ymin), (Xmax, Ymax) coordinates of the objects.

In my project, I have data set where every image has single character only

How to create the bounding box where all the characters are already in Region of Interest...

Thanks

Answer 1

LabelImg is the most commonly used tool for creating bounding boxes on images.
Also, you said that all characters are within the region of interest, one way to approach this is to create a new image containing all of these characters then manually adding the bounding boxes for each.

It is also worth noting that Object Detection algorithms also consider the surroundings containing the said object, thus more variance in the environment that includes desired objects in training will result in a more robust model when exposed to a foreign/new environment.

Also, you should also consider using data augmentation like rotation, scaling, and translation. Because these parameters are also encountered in real-time detection thus also contributes for a more robust model that can detect accurately in varying surroundings.

You can read more about Tensor Flow Data Augmentation in this link.