PaperSummary09 : PolarMask++

The paper proposes PolarMask++, a framework that uses polar representation for efficient single-shot instance segmentation. It reformulates the task as contour prediction in polar coordinates (angles and distances from a center point). It unifies instance segmentation and object detection within a single framework without depending on bounding box predictions.

The methodology is:

1. PolarMask++ Framework: It combines a backbone network, refined feature pyramid and task-specific heads for classification, centerness prediction and mask regression. Additionally, it introduces soft polar centerness and Polar IoU loss to improve training efficiency and segmentation accuracy.
2. Refined feature pyramid enhances feature representation across scales, benefiting small object segmentation.
3. Polar IoU loss optimizes mask regression by treating the rays collectively rather than independently.

PolarMask++ is computationally efficient and demonstrates state of the art perfromance on benchmark datasets. It is beneficial in speed, simplicity and scalability.

References:

PolarMask++: Enhanced Polar Representation for Single-Shot Instance Segmentation and Beyond