Follow and Star
No More Getting Lost
Computer Vision Research Institute
Public Account ID｜ComputerVisionGzq
Study Group｜Scan the QR code on the homepage to get the joining method
Computer Vision Research Institute Column
Object detection is considered one of the most challenging problems in the field of computer vision, as it involves the combination of object classification and object localization in a scene. Today, we share this framework which may be a bit outdated, but it captures the essence!
1. Introduction
Object detection is regarded as one of the most challenging problems in the field of computer vision, as it involves the combination of object classification and object localization in a scene. Recently, deep neural networks (DNNs) have been shown to achieve outstanding object detection performance compared to other methods, with YOLOv2 being one of the state-of-the-art techniques based on DNNs.
Object detection methods are evaluated based on speed and accuracy. Although YOLOv2 can achieve real-time performance on powerful GPUs, utilizing this method for real-time object detection in videos on embedded computing devices with limited computational power and memory remains very challenging.
2. Overview
In today’s sharing, researchers proposed a new framework called Fast YOLO, which is a fast You Only Look Once framework that accelerates YOLOv2 to enable real-time object detection in videos on embedded devices.
First, an evolutionary deep intelligence framework is used to evolve the YOLOv2 network architecture and produce an optimized architecture (referred to as O-YOLOv2), whichreduces parameters by 2.8 times and decreases IOU by about 2%. To further reduce power consumption on embedded devices while maintaining performance, a motion-adaptive inference method is introduced in the proposed Fast YOLO framework to reduce the frequency of deep inference based on temporal motion characteristics of O-YOLOv2. Experimental results show that compared to the original YOLOv2, the proposed Fast YOLO framework can reduce the number of deep inferences by an average of 38.13% and achieve an average speedup of about 3.3 times for object detection in videos, resulting in Fast YOLO running at an average of about 18 FPS on the Nvidia Jetson TX1 embedded system.
3. New Framework
The proposed Fast YOLO framework consists of two main parts: i) the optimized YOLOv2 architecture, and ii) motion-adaptive inference (see the above image). For each video frame, an image stack composed of video frames with a reference frame is passed to a 1×1 convolution layer. The result of the convolution layer is a motion probability map, which is then sent to the motion-adaptive inference module to determine whether deep inference is needed to compute the updated class probability map. As mentioned in the introduction, the main goal is to introduce a framework for object detection in videos that can execute faster on embedded devices while reducing resource usage, thereby significantly lowering power consumption. By utilizing this motion-adaptive inference method, the frequency of deep inference is greatly reduced and executed only when necessary.
One of the main challenges of deep neural networks, especially when used in embedded scenarios, is the design of the network architecture. The design process is typically performed by human experts who explore a large number of network configurations to find the optimal architecture in terms of modeling accuracy and the number of parameters for a specific task. Finding an optimized network architecture is currently often treated as a hyperparameter optimization problem, but this approach is very time-consuming, and most methods are either computationally intractable for large network architectures or lead to suboptimal solutions that are not suitable for embedded use.
For example, a common method for hyperparameter optimization is grid search, which examines a large number of different network configurations and then selects the best configuration as the final network architecture. However, deep neural networks designed for object detection in videos (such as YOLOv2) have a large number of parameters, making it computationally difficult to search the entire parameter space for the optimal solution.
Therefore, the researchers did not utilize hyperparameter optimization methods to obtain the optimal network architecture based on YOLOv2, but instead employed network optimization strategies specifically designed to improve network efficiency. In particular, the researchers utilized an evolutionary deep intelligence framework to optimize the network architecture to synthesize deep neural networks that meet the memory and computational power constraints of embedded devices.
To further reduce the power consumption of processing units for embedded object detection in videos, the researchers leveraged the fact that not all captured video frames contain unique information, and therefore deep inference is not required for all frames. Thus, the researchers introduced a motion-adaptive inference method to determine whether a specific video frame requires deep inference. By using the previously introduced O-YOLOv2 network for deep inference only when necessary, this motion-adaptive inference technique helps the framework reduce the demand for computational resources, significantly lowering power consumption and improving processing speed.
4. Experiments
Comparison of architecture and performance between the original YOLOv2 network architecture and the optimized YOLOv2
Average runtime performance and deep inference frequency of the proposed Fast YOLO, O-YOLOv2, and original YOLOv2 running on the Nvidia Jetson TX1 embedded system.
END
For reprints, please contact this public account for authorization
The Computer Vision Research Institute study group is waiting for you to join!
ABOUT
Computer Vision Research Institute
The Computer Vision Research Institute mainly involves the field of deep learning, focusing on research directions such as object detection, object tracking, image segmentation, OCR, model quantization, and model deployment. The institute shares the latest paper algorithms and new frameworks daily, provides one-click downloads of papers, and shares practical projects. The institute emphasizes “technical research” and “practical implementation.” It will share practical processes in different fields, allowing everyone to truly experience real scenarios beyond theory, cultivating the habit of hands-on programming and thoughtful consideration!
Previous Recommendations
🔗
- Drone-YOLO: An Effective Object Detection for Drone Images
- Airport Project: Solving Real-Time Object Detection Issues Caused by Flight Object Size/Aspect Ratio, Speed, Occlusion, etc.
- 2PCNet: Day-Night Unsupervised Domain Adaptation for Object Detection (with Original Code)
- YOLO-S: A Lightweight and Accurate Class YOLO Network for Small Object Detection
- Major Revision of Yolo Framework | A New Framework for Object Detection with Extremely Low Energy Consumption (Paper Download Included)
- Improved Detection Algorithm: For Object Detection in High-Resolution Optical Remote Sensing Images
- Xiaomi Tablet 6 Max-Yolo: Real-Time Detection on Portable Devices is No Longer a Challenge
- The Importance of Rotational Angle Object Detection!!! (with Source Paper Download)