Possible issues:

• Too few epochs, 20 is too low to a decent convergence. Leave it to the default, or start with at least 100. You can even put a higher number like 1k and set some early stopping strategy.

• 80 images is also kind of low. Try to increase it to at least 1k. You can synthetically increase it by using data augmentation, such as Albumentation library. There is a way to synthesize images during yolo training, please take a look on Yolo docs for configuring this properly. I would use mostly the lightning, contrast, rotation, translation, crop and eraser functions.

• If your input image is large, specially the one taken from distance, you might get better accuracy working on a sliding window. The easiest way might be SAHI (https://docs.ultralytics.com/guides/sahi-tiled-inference/)