Instance Segmentation

Instance Segmentation is a subfield of computer vision that involves identifying each individual object of interest within an image and delineating its specific boundaries. It is more granular than semantic segmentation, which labels pixels belonging to the same class with the same color but doesn't differentiate between individual objects of the same class. In instance segmentation, each object, even of the same class, is separately identified and marked.

How Instance Segmentation works

Instance segmentation works by classifying every pixel in an image not only to a specific label but also to a unique object. This is usually done using deep learning models. There are two main approaches to achieving instance segmentation: "top-down" and "bottom-up".

In a top-down approach, using models like Mask R-CNN, the model first detects objects in the image and then segments each detected object. Detection is performed by drawing bounding boxes around detected objects, and then within each bounding box, a mask is created for the detected object.

In a bottom-up approach, using methods like discriminative loss function, pixels or superpixels are first grouped together and then these groups are associated together to form objects.

Regardless of the approach, the output of instance segmentation is a binary mask for each detected object in the image, distinguishing not only different classes of objects, but also different instances within the same class. For example, in an image with several cars, each car will be separately identified and labeled.