# DAMO-YOLO : A Report on Real-Time Object Detection Design

**Authors:** Xianzhe Xu, Yiqi Jiang, Weihua Chen, Yilun Huang, Yuan Zhang, Xiuyu Sun  
**arXiv:** [2211.15444](https://arxiv.org/abs/2211.15444)  
**Date:** 2022-11-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.15444v4/

## Key points

- Authors introduce novel object detection method called DAMO-YOLO
- Incorporates cutting-edge technologies such as Neural Architecture Search (NAS), Reparameterized Generalized-FPN, lightweight head with AlignedOTA label assignment, and distillation enhancement
- Optimization of detection backbone using MAE-NAS guided by principle of maximum entropy
- Structures resembling ResNet/CSP with spatial pyramid pooling and focus modules
- Integration of Generalized-FPN with accelerated queen-fusion for detector neck, enhanced CSPNet with efficient layer aggregation networks (ELAN) and reparameterization
- Study on detector head size impact on accuracy, favoring heavy neck with single task projection layer
- Introduction of AlignedOTA to address misalignment issues in label assignment, distillation schema for performance enhancement
- Development of range of models tailored to different scenarios: DAMO-YOLO-T/S/M/L for general industry requirements achieving mAPs of 43.6/47.7/50.2/51.9 on COCO dataset with latencies ranging from 2.78 to 7.95 ms on T4 GPUs; DAMO-YOLO-Ns/Nm/Nl lightweight models for edge devices achieving mAPs of 32.3/38.2/40.5 on COCO with latencies between 4.08 and 6.69 ms on X86-CPU
- Outperforms existing YOLO series models in various application scenarios due to innovative technologies and scalable model designs tailored to specific needs

## Summary

In their report titled "DAMO-YOLO: A Report on Real-Time Object Detection Design," authors Xianzhe Xu, Yiqi Jiang, Weihua Chen, Yilun Huang, Yuan Zhang, and Xiuyu Sun introduce a novel object detection method called DAMO-YOLO. This method surpasses the performance of the well-known YOLO series by incorporating cutting-edge technologies such as Neural Architecture Search (NAS), efficient Reparameterized Generalized-FPN (RepGFPN), a lightweight head with AlignedOTA label assignment, and distillation enhancement. The authors employ MAE-NAS guided by the principle of maximum entropy to optimize the detection backbone for low latency and high performance. This results in structures resembling ResNet/CSP with spatial pyramid pooling and focus modules. Following the design philosophy of "large neck, small head," they integrate Generalized-FPN with accelerated queen-fusion for the detector neck and enhance CSPNet with efficient layer aggregation networks (ELAN) and reparameterization. Furthermore, the study explores how detector head size impacts detection accuracy, concluding that a heavy neck with a single task projection layer yields superior results. The introduction of AlignedOTA addresses misalignment issues in label assignment, while a distillation schema enhances overall performance. Based on these advancements, the authors develop a range of models tailored to different scenarios. For general industry requirements, they propose DAMO-YOLO-T/S/M/L models achieving mAPs of 43.6/47.7/50.2/51.9 on COCO dataset with latencies ranging from 2.78 to 7.95 ms on T4 GPUs. Additionally, for edge devices with limited computing power, they introduce DAMO-YOLO-Ns/Nm/Nl lightweight models achieving mAPs of 32.3/38.2/40.5 on COCO with latencies between 4.08 and 6.69 ms on X86-CPU. Overall, the proposed DAMO-YOLO method outperforms existing YOLO series models in various application scenarios due to its innovative technologies and scalable model designs tailored to specific needs.