Instance Segmentation Via Instance Categories: A Simple And Flexible Framework.pdf

Instance segmentation is challenging due to varying numbers of instances and the need for pixel-level semantic labeling. Existing methods fall into two categories: 'detect-then-segment' (e.g., Mask R-CNN) or bottom-up approaches using embedding vectors and clustering.

SOLO takes a different perspective by introducing "instance categories" based on an object's location and size, effectively converting instance mask segmentation into a classification problem. The key components are:

1. Locations: An image is divided into a grid of SxS cells, with each cell representing a potential center location for an object. Each output channel in the network corresponds to a specific location category, allowing direct prediction of instance masks based on pixel-wise classification.

2. Sizes: A Feature Pyramid Network (FPN) is used to handle objects of varying sizes by assigning them to different levels of feature maps, enabling regular separation and classification by "instance categories."

Advantages:

Simplicity: SOLO optimizes the network end-to-end using only mask annotations, avoiding the need for box detection or pixel grouping.
Performance: Achieves on par accuracy with Mask R-CNN on the challenging MS COCO dataset and outperforms other single-shot instance segmenters.
Flexibility: The framework is generalizable to tasks like instance contour detection with minimal modifications.

Key Contributions:

Converts coordinate regression into classification through discrete quantization of location and size features.
* Demonstrates the potential for leveraging advances in semantic segmentation techniques for instance segmentation.