Custom data images

[albertofernandezvillan]

I am creating some images to train a custom object detector with YOLO V4.
For example, suppose I want to train an object detector for detecting bicycles in thermal images, and also, I want to create the minimal number of images for training/testing/validating. For this purpose, I am performing "artificial images" with objects over a real background. See the two attached images as an example.
The question is: Is it a good idea to have this kind of images or is it a bad idea?
Many thanks in advance

[stephanecharette]

Note that darknet already has built-in support for mosaic. You simply have to turn it on in the .cfg file. For example: https://www.ccoderun.ca/darkmark/DataAugmentationMisc.html#Mosaic

[albertofernandezvillan]

Thank you for the answer. I am aware of this feature. This is why I have asked it. If I am not wrong mosaic data augmentation combines 4 images into one and different scales and ratios allowing the model to learn how to identify objects at different scales.
Therefore, I don't know if creating the images for training that I have introduced in this question is a good idea or not. I suppose it is not bad, but I am looking for feedback from other users.