Matching Strategies
Introduction
This package provides implementations of various matching strategies, all of which are particular cases of Optimal Transport. These strategies are crucial in object detection tasks, as they determine how predictions are matched with ground truth objects. The module provides a range of strategies, including ClosestTarget (used in methods like Faster R-CNN and SSD), Hungarian (used in DETR and its extensions) and UnbalancedSinkhorn (which generalizes the other strategies). Each strategy is implemented as a separate class, allowing for easy extension and customization.