# Mask R-CNN

**Authors:** Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick  
**arXiv:** [1703.06870](https://arxiv.org/abs/1703.06870)  
**Date:** 2017-03-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1703.06870v1/

## Key points

- Authors Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick introduce Mask R-CNN for object instance segmentation
- Mask R-CNN efficiently detects objects in images and generates high-quality segmentation masks
- Method builds upon Faster R-CNN by adding a branch for predicting object masks alongside bounding box recognition
- Ease of training and minimal overhead allow Mask R-CNN to run at 5 frames per second
- Versatile framework can be adapted to tasks beyond instance segmentation, such as human pose estimation
- Achieves top results in COCO suite challenges without specialized techniques or tricks
- Surpasses existing single-model entries and outperforms winners of the COCO 2016 challenge
- Authors aim for their approach to be a solid baseline in instance-level recognition and plan to share their code for further research.

## Summary

In their technical report titled "Mask R-CNN," authors Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick introduce a novel framework for object instance segmentation that is conceptually simple, flexible, and general. The proposed approach efficiently detects objects within an image while simultaneously generating high-quality segmentation masks for each instance. Referred to as Mask R-CNN, this method builds upon the Faster R-CNN architecture by incorporating a branch dedicated to predicting object masks in parallel with the existing branch for bounding box recognition. One of the key advantages of Mask R-CNN is its ease of training and minimal overhead on top of Faster R-CNN, allowing it to run at an impressive speed of 5 frames per second. Additionally, the framework's versatility enables straightforward adaptation to various tasks beyond instance segmentation; for example, it can be utilized for estimating human poses within the same model structure. The authors demonstrate the effectiveness of Mask R-CNN by achieving top results across all three tracks of the COCO suite of challenges: instance segmentation, bounding-box object detection, and person keypoint detection. Notably, without employing any specialized techniques or "tricks," Mask R-CNN surpasses all existing single-model entries on every task and outperforms even the winners of the COCO 2016 challenge. Overall, the authors aim for their straightforward yet powerful approach to serve as a solid baseline in the field of instance-level recognition and plan to make their code available to facilitate further research and development in this area.