I've managed to make it work. You just provide your buffered polygons as your Input Feature Class and don't use Input Mask Polygons.

Hello, thanks for proposing, it seems like the way to go. I've prepared the buffered dataset as you suggested however I cannot get the Export Training Data to work as I intend. I want each buffered bounding box to be exported into one image and a metadata file containing the object class and the box coordinates. However, it seems to ignore input_mask_polygons and generate chips based on the Tile Size X and Tile Size Y. Can you suggest proper settings for the Export Training Data tool?