Techniques from image processing, computer vision, and computer graphics can be borrowed to improve the quality of the output as well as detection. As an example, a typical image processing task is performing the morphology operations of opening and closing. This combination is commonly used to remove noise and fill in small holes in a binary image. Another useful post-processing task we could borrow from computer vision is watersheds, a segmentation technique that treats the image as a topographical map, where the intensity of change defines the ridges and the boundary of the fill (or segmentation).

Another tool to use for post-processing is another model. You're familiar with YOLO for object detection. We can apply it to obtain its predicted boundaries of the object, which we can then use to refine our segmentation. Another model, and one being adopted for this task, is conditional random fields (CRF), which is capable of smoothing out the edges of our mask. 

There are a vast number of techniques available from the fields of image processing, computer vision, and computer graphics, and I strongly encourage you to explore each area to build up your tool set.