EdgeYOLO: An Edge-Real-Time Object Detector (2302.07483v1)

Abstract: This paper proposes an efficient, low-complexity and anchor-free object detector based on the state-of-the-art YOLO framework, which can be implemented in real time on edge computing platforms. We develop an enhanced data augmentation method to effectively suppress overfitting during training, and design a hybrid random loss function to improve the detection accuracy of small objects. Inspired by FCOS, a lighter and more efficient decoupled head is proposed, and its inference speed can be improved with little loss of precision. Our baseline model can reach the accuracy of 50.6% AP50:95 and 69.8% AP50 in MS COCO2017 dataset, 26.4% AP50:95 and 44.8% AP50 in VisDrone2019-DET dataset, and it meets real-time requirements (FPS>=30) on edge-computing device Nvidia Jetson AGX Xavier. We also designed lighter models with less parameters for edge computing devices with lower computing power, which also show better performances. Our source code, hyper-parameters and model weights are all available at https://github.com/LSH9832/edgeyolo.

EdgeYOLO: An Edge-Real-Time Object Detector

The paper "EdgeYOLO: An Edge-Real-Time Object Detector" presents an innovative approach to addressing the challenges of object detection on edge computing platforms. This work is built on the YOLO framework, widely recognized for its efficiency and speed in object detection tasks.

Core Contributions

The authors introduce an efficient anchor-free object detector tailored for edge devices. Key contributions include:

• Enhanced Data Augmentation: A novel data augmentation technique is proposed to mitigate overfitting and enhance small object detection efficacy. This involves a flexible combination of Mosaic and Mixup methods, ensuring rich and valid data input while maintaining effective labels.
• Hybrid Random Loss Function: The paper introduces a loss function designed to improve small object detection accuracy, contributing to increased overall detection performance.
• Lite-Decoupled Head: Inspired by the FCOS architecture, a streamlined decoupled head is incorporated to balance inference speed with detection precision.
• Model Reduction Techniques: These techniques optimize the computational demands, enabling the model to achieve real-time processing speeds on edge devices like the Nvidia Jetson AGX Xavier without significant precision loss.

Performance and Results

EdgeYOLO demonstrates notable results in benchmark datasets:

• On the MS COCO2017 dataset, it achieves 50.6% AP$_{50:95}$ and 69.8% AP$_{50}$, successfully maintaining performance criteria for real-time applications (FPS $\geq$ 30 on the Nvidia Jetson AGX Xavier).
• On the VisDrone2019-DET dataset, the model achieves 26.4% AP$_{50:95}$ and 44.8% AP$_{50}$.

The model's emphasis on reducing inference latency and retaining high frame rates on edge devices underscores its practical application potential.

Technical Details

1. Anchor-Free Architecture: The decision to leverage an anchor-free approach results in reduced computational complexity, ideal for edge devices where processing power and energy efficiency are constrained.
2. Data Augmentation: By mixing strategies like enhanced-Mosaic and Mixup, the model effectively enriches training data, addressing overfitting and improving robustness in small object detection.
3. Loss Function Optimization: The staged training with hybrid random loss shows a sophisticated attempt to balance small object precision with overall detection quality. This balance is crucial in applications requiring real-time detection without sacrificing accuracy.
4. Lightweight Decoupled Head: By optimizing the decoupled head, the paper ensures that the increase in speed does not compromise precision, a common issue in real-time object detection frameworks.

Theoretical and Practical Implications

This research has significant implications for the deployment of real-time AI models in resource-constrained environments. The methodological innovations, particularly in data augmentation and loss optimization, provide insights for future development of edge-compatible AI systems. Moreover, the practical application potential of EdgeYOLO extends to autonomous systems and mobile devices, where real-time object detection is critical.

Future Directions

Further exploration might focus on enhancing the accuracy of small object detection, perhaps through advanced feature pyramid networks or integrating semantic information. Additionally, extending the framework to other real-time tasks such as instance segmentation could broaden its applicability.

In summary, EdgeYOLO offers a compelling solution for edge-based object detection, balancing precision with the practical constraints of edge computing environments. The paper's methodologies and results should inspire ongoing research initiatives in the field of efficient and effective edge AI applications.