# SOLO: A Simple Framework for Instance Segmentation

**Authors:** Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, Lei Li  
**arXiv:** [2106.15947](https://arxiv.org/abs/2106.15947)  
**Date:** 2021-06-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.15947v1/

## Key points

- The paper introduces the concept of "instance categories" for instance segmentation
- Existing approaches for instance segmentation either follow a 'detect-then-segment' strategy or predict embedding vectors and cluster pixels into individual instances
- SOLO (Segmenting Objects by Locations) is a direct and fast framework for instance segmentation that assigns categories to each pixel based on its location
- SOLO eliminates the need for grouping post-processing or bounding box detection
- Several variants of SOLO are derived, including Vanilla SOLO, Decoupled SOLO, and Dynamic SOLO
- SOLO achieves state-of-the-art results in terms of both speed and accuracy while being simpler than existing methods
- Besides instance segmentation, SOLO also yields impressive results in object detection and panoptic segmentation
- SOLO can be extended to perform one-stage instance-level image matting

## Summary

The paper titled "SOLO: A Simple Framework for Instance Segmentation" introduces a novel perspective on instance segmentation by introducing the concept of "instance categories". Unlike other dense prediction tasks, instance segmentation is challenging due to the arbitrary number of instances. Existing approaches either follow a 'detect-then-segment' strategy or predict embedding vectors and cluster pixels into individual instances. In this paper, the authors propose SOLO (Segmenting Objects by Locations), a direct and fast framework for instance segmentation. The key idea behind SOLO is to assign categories to each pixel within an instance based on its location. This approach eliminates the need for grouping post-processing or bounding box detection. The authors derive several variants of SOLO, including Vanilla SOLO, Decoupled SOLO, and Dynamic SOLO, all following the same principle. By directly mapping a raw input image to object categories and instance masks, SOLO achieves state-of-the-art results in terms of both speed and accuracy while being simpler than existing methods. Additionally, besides instance segmentation, SOLO also yields impressive results in object detection and panoptic segmentation. The authors further demonstrate the flexibility and high-quality segmentation of SOLO by extending it to perform one-stage instance-level image matting. Overall, this paper presents a novel approach that outperforms existing methods in terms of simplicity and performance.