YOLACT

YOLACT, which stands for "You Only Look At Coefficients," is a real-time instance segmentation algorithm. It was developed to address the challenge of performing instance segmentation, the task of identifying and delineating each distinct object instance within an image, at a speed comparable to object detection. Traditional instance segmentation methods often involve multiple stages, which can be computationally expensive and slow. YOLACT aims to unify these stages into a single, efficient network.

The core idea behind YOLACT is to decouple the segmentation mask generation from the object localization. Instead of predicting masks directly for each detected object, YOLACT generates a set of prototype masks and then learns a linear combination of these prototypes to produce the final instance masks. The coefficients for this combination are predicted along with the bounding boxes and class predictions, hence the name "You Only Look At Coefficients." This approach allows for a more parallel and faster computation.

YOLACT's architecture consists of two main branches. The first branch predicts a set of prototype masks that are generally applicable across the entire image. The second branch predicts per-instance bounding box coordinates, class probabilities, and the coefficients that will be used to combine the prototype masks to form the final instance-specific masks. This efficient design enables YOLACT to achieve real-time performance, making it suitable for applications where speed is critical, such as autonomous driving and video analysis.