I am in the process of training a custom Model to detect laser pointers pointed on a white canvas. A camera will be pointed on the opposite side if the canvas. The goal is to count the laser dots and detect if one is missing (meaning one beam has been interrupted).
As I am completely new to the topic of object detection I have a few questions on how to approach this:
- Will it make a difference if I train the model with pictures containing only one laser dot, or with pictures containing multiple dots? In the final application there will be 6-24 dots present.
- I think of using a fisheye camera to better capture the area (canvas is about 6 meters wide). In this case, is it better to
a) Train the model with pictures shot with a normal camera and convert the fisheye images on the fly to panoramic before analyzing them (like here: How to use YOLOv5 with fisheye camera - YouTube) or
b) Train the model with images captured with the fisheye lens and not convert them before analyzing?
In my case, it does not matter if the model gets the bounding box of the dots right, the main goal is to count the dots. It will also not be necessary to augment the image, I really only look for a number as an output.
Thank you all in advance for any help and suggestions.