Enabling Computer Vision: Object Detection with StarFive VisionFive 2 using GPU Acceleration
Tip:
If you are interested in these applications, join us at the embedded world China Conference on June 14 to 16, or stay tuned to our application center for more resources.
Recently, StarFive has implemented a GPU-accelerated object detection application on VisionFive 2 by using the OpenCV framework and Debian OS.
Object detection is widely used in various applications, including video surveillance, autonomous driving, face recognition, object recognition, and pedestrian detection. OpenCV, as a powerful computer vision library, provides a rich set of functions and tools for processing image and video data, enabling various computer vision tasks and making object detection easier and more efficient.
OpenCV User Application
This application:
- Supports deploying object detection applications using languages such as Python and C++;
- Supports multiple DNNs (deep neural network interfaces)
- Allows GPU acceleration with OpenCL. Additionally, the object detection functionality on VisionFive 2 supports video input from CSI and USB cameras and displays the detection results on an HDMI screen with QT5 as the GUI backend.
Application Highlights:
- Based on the Debian OS & OpenCV framework.
- Utilizing GPU acceleration through OpenCL and customized operator optimizations for VisionFive 2. The DNN model inference speed is 1.5 to 2.5 times faster than using CPU computation on a Raspberry Pi 3B.
- Supports inference for various popular deep learning framework models such as Darknet & Caffe, as well as the universal ONNX format.
- Provides application routines including versions in both Python and C++ languages, facilitating easy adoption and secondary development.
- Supports a wide range of video input and output connections, including the Raspberry Pi Camera module 2, USB cameras, HDMI displays, and more.
The specific applications supported by our object detection solution on VisionFive 2 are as follows: - QR code detection and decoding application;
- Universal object detection based on MobileNet SSD;
- Universal object detection based on YOLO-V5 ONNX format model;
- Object detection based on the YOLO-V3 model;
- Image edge detection;
- Defect detection;
Application Scenarios:
- Industrial computer vision field
- Deep learning academic research and student experimental development
- Visual modules for commercial products
Demos
This project will be released with the new Debian system in the form of deb packages later.