Fast YOLO: A Fast You Only Look Once System for Real-time Embedded Object Detection in Video (1709.05943v1)

Abstract: Object detection is considered one of the most challenging problems in this field of computer vision, as it involves the combination of object classification and object localization within a scene. Recently, deep neural networks (DNNs) have been demonstrated to achieve superior object detection performance compared to other approaches, with YOLOv2 (an improved You Only Look Once model) being one of the state-of-the-art in DNN-based object detection methods in terms of both speed and accuracy. Although YOLOv2 can achieve real-time performance on a powerful GPU, it still remains very challenging for leveraging this approach for real-time object detection in video on embedded computing devices with limited computational power and limited memory. In this paper, we propose a new framework called Fast YOLO, a fast You Only Look Once framework which accelerates YOLOv2 to be able to perform object detection in video on embedded devices in a real-time manner. First, we leverage the evolutionary deep intelligence framework to evolve the YOLOv2 network architecture and produce an optimized architecture (referred to as O-YOLOv2 here) that has 2.8X fewer parameters with just a ~2% IOU drop. To further reduce power consumption on embedded devices while maintaining performance, a motion-adaptive inference method is introduced into the proposed Fast YOLO framework to reduce the frequency of deep inference with O-YOLOv2 based on temporal motion characteristics. Experimental results show that the proposed Fast YOLO framework can reduce the number of deep inferences by an average of 38.13%, and an average speedup of ~3.3X for objection detection in video compared to the original YOLOv2, leading Fast YOLO to run an average of ~18FPS on a Nvidia Jetson TX1 embedded system.

Overview of "Fast YOLO: A Fast You Only Look Once System for Real-time Embedded Object Detection in Video"

The paper presents Fast YOLO, an enhancement of the YOLOv2 architecture, aimed at addressing constraints in performing object detection on embedded systems. Fast YOLO achieves real-time video processing on devices with limited computational resources, such as smart phones and surveillance cameras.

Key Contributions

The paper introduces two primary strategies in developing Fast YOLO:

1. Optimized YOLOv2 Architecture: Utilizing the evolutionary deep intelligence framework, the authors evolved the network architecture of YOLOv2 to create a streamlined version designated as O-YOLOv2. This optimized architecture exhibits a 2.8-fold reduction in parameters with only a slight decrease (~2%) in Intersection over Union (IOU), a measure of detection accuracy. The reduced parameter set translates to faster inference and lower power consumption, making it suitable for real-time applications in resource-constrained environments.
2. Motion-Adaptive Inference: To further enhance efficiency, a motion-adaptive inference method was integrated. This technique assesses temporal motion to decide when full inference is necessary, minimizing the frequency and computational expense of deep inference processes. By doing so, Fast YOLO significantly decreases the number of deep inferences required, reporting an impressive reduction of approximately 38.13%.

Experimental Results

The effectiveness of Fast YOLO is substantiated through experimental evaluations on the Nvidia Jetson TX1 embedded system. Here, Fast YOLO demonstrated a speed-up of approximately 3.3 times over the original YOLOv2, achieving an average of 18 frames per second. These results highlight both the practical benefits in terms of speed and the strategic balance in maintaining detection accuracy despite architectural simplifications.

Implications and Future Directions

The advancements presented in Fast YOLO have noteworthy practical implications for embedded systems in real-time applications such as autonomous vehicles and portable devices. By achieving efficient object detection with constrained resources, Fast YOLO demonstrates potential in enabling edge-based intelligence rather than relying on cloud servers, thus enabling faster and more economically viable real-world implementations.

Theoretically, Fast YOLO opens avenues for further exploration into the synthesis of deep neural networks, specifically in optimizing architectures without significant sacrifices in performance metrics. Future research could assess the application of similar evolutionary frameworks in developing architectures suited for other tasks or domains, perhaps extending beyond video object detection.

In addition, the motion-adaptive inference module presents an interesting aspect of dynamically modulating the inference pipeline based on environmental and contextual understanding. Future work might explore sophisticated models capable of even more nuanced detections of relevance or importance, perhaps through integrating more advanced predictive or context-aware capabilities.

Overall, this paper marks a significant contribution to the field, particularly within the constraints of real-world embedded system environments, offering a framework that adeptly balances speed, resource consumption, and accuracy.