Add background replacement technique for COCO Instance Segmentation

[juanmed]

Implementation of the background replacement technique for the COCO Instance Segmentation problem. This works by giving an input directory where background images are stored, creating a background_replacement technique and passing the input directory in the parameters dictionary. The technique will:

• Load and resize a random background image to fit the image to augment
• Create a mask of all objects instances
• Replace all pixels that do not correspond to instances with the corresponding pixel (same position) from the background image.
• The labels and masks remain unchanged.

[joheras]

Thanks for your work. This is a nice improvement for the library. Best, Jónathan