training loss remains constant even though CNN, optimizers etec are changed

Question

I am trying to train a CNN with OSU RGB Thermal dataset to predict bounding boxes for pedestrians. I have collected images with only one person walking in the frame and compiled