# OMG-Seg: Is One Model Good Enough For All Segmentation?

**Authors:** Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, Chen Change Loy  
**arXiv:** [2401.10229](https://arxiv.org/abs/2401.10229)  
**Date:** 2024-01-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.10229v1/

## Key points

- OMG-Seg is a novel approach to handle multiple segmentation tasks within a single model
- It can handle image semantic segmentation, instance segmentation, panoptic segmentation, and their video counterparts
- It can also handle open vocabulary settings, prompt-driven interactive segmentation, and video object segmentation
- OMG-Seg is the first model to successfully address all these diverse tasks while achieving satisfactory performance
- The transformer-based encoder-decoder architecture with task-specific queries and outputs supports over ten distinct segmentation tasks
- It reduces computational and parameter overhead across various datasets
- Inter-task influences and correlations are rigorously evaluated during co-training for optimal performance
- Code and models for OMG-Seg are available on the authors' GitHub repository at https://github.com/lxtGH/OMG-Seg

## Summary

, , , , 

In their paper "OMG-Seg: Is One Model Good Enough For All Segmentation? ", Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy propose a novel approach called OMG-Seg to efficiently and effectively handle multiple segmentation tasks within a single model. They demonstrate that this unified framework can handle image semantic segmentation, instance segmentation, panoptic segmentation, as well as their video counterparts. Additionally, it can also handle open vocabulary settings, prompt-driven interactive segmentation like SAM (Scribble-And-Mask), and video object segmentation. This makes OMG-Seg the first model to successfully address all these diverse tasks while achieving satisfactory performance. The authors show that this transformer-based encoder-decoder architecture with task-specific queries and outputs can support over ten distinct segmentation tasks while reducing computational and parameter overhead across various datasets. They also rigorously evaluate the inter-task influences and correlations during co-training to ensure optimal performance. To facilitate further research and implementation of OMG-Seg, the authors provide code and models on their GitHub repository at https://github.com/lxtGH/OMG-Seg. Overall,<kgd>OMG-Seg</kgd> presents a comprehensive solution for handling multiple segmentation tasks with improved efficiency and effectiveness through its unified framework.