# MatrixNets: A New Scale and Aspect Ratio Aware Architecture for Object Detection

**Authors:** Abdullah Rashwan, Rishav Agarwal, Agastya Kalra, Pascal Poupart  
**arXiv:** [2001.03194](https://arxiv.org/abs/2001.03194)  
**Date:** 2020-01-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2001.03194v1/

## Key points

- MatrixNets (xNets) is a deep architecture designed for object detection
- xNets enable scale and aspect ratio aware architecture by mapping objects with similar sizes and aspect ratios into specialized layers
- xNets enhance single-stage object detection frameworks in anchor-based object detection by accurately localizing objects in images
- xNets are also utilized in corner-based object detection by predicting the top-left and bottom-right corners of objects, while each corner predicts the center location of the respective object
- The proposed architecture achieved a mean Average Precision (mAP) score of 47.8% on MS COCO dataset, surpassing its CornerNet counterpart by +5.6 mAP
- This architecture bridges the gap between single-stage and two-stage detectors, providing an effective solution for object detection tasks with scale and aspect ratio awareness

## Summary

The paper presents MatrixNets (xNets), a novel deep architecture designed for object detection. xNets are able to map objects with similar sizes and aspect ratios into specialized layers, enabling the network to have a scale and aspect ratio aware architecture. This is achieved by leveraging xNets to enhance single-stage object detection frameworks. The authors first apply xNets on anchor-based object detection, where they predict the centers of objects and regress the top-left and bottom-right corners. By doing so, they are able to accurately localize objects in images. Next, they utilize MatrixNets for corner-based object detection by predicting the top-left and bottom-right corners of objects. Interestingly, each corner also predicts the center location of the respective object. This approach enhances corner-based detection by replacing the embedding layer with center regression. To evaluate their proposed architecture, the authors conducted experiments on MS COCO dataset. The final architecture achieved a mean Average Precision (mAP) score of 47.8%, surpassing its CornerNet counterpart by +5.6 mAP. Additionally, this architecture helps bridge the gap between single-stage and two-stage detectors, providing an effective solution for object detection tasks by incorporating scale and aspect ratio awareness into their deep architecture design.