Nowadays AI computing is driving many
traditional industries towards intelligent. No matter in security, industrial
quality inspection, ADAS (Advanced Driver Assistance System), retail or many
other industries, AI computing & technology is unfolding a new era. In the
future of human life, AI technology will be everywhere.

FZ3 Card is a deep learning accelerator card produced by MYIR while cooperating with Baidu. One the hardware side, the FZ3 Card is built based on Xilinx Zynq UltraScale+ ZU3EG MPSoC which integrates 64-bit quad-core ARM Cortex-A53, GPU and FPGA, thus comes with multi-core processing capability, FPGA programmable capability and hardware decoding capability for video stream. On the software side, the FZ3 Card has built-in deep learning soft core based on Linux OS and Baidu PaddlePaddle deep learning AI (Artificial Intelligence) framework which is fully compatible to use Baidu Brain’s model resources and AI development tools like EasyDL, AI Studio and EasyEdge to enable developers and engineers to quickly train-deploy-reasoning models. Provided with these hardware capabilities and software resources, the FZ3 Card reduces the threshold of development validation, product integration, scientific research and teaching significantly.
FZ3 Card comes with Linux OS, users may develop applications based on Linux OS. Detailed steps have been introduced in this page before: http://www.myirtech.com/news_list.asp?id=827.
Main procedures are as below:
1. Obtaining models
2. Connecting video data source
3. Load device driver
4. Using the prediction library
5. Create applications
The system block for the soft core is show as below:
Performance data for the
common models (untailored) on FZ3 Card:
Network
|
Pixels
|
Single Frame time consuming
(ms)
|
resnet50
|
224 x 224
|
42ms
|
mobilenet-v1
|
224 x 224
|
10ms
|
inception-v2
|
299 x 299
|
41ms
|
inception-v3
|
299 x 299
|
70ms
|
resnext
|
224×224
|
69ms
|
mobilenet-ssd
|
300 x 300
|
24ms
|
mobilenet-ssd-640
|
640 x 640
|
79ms
|
vgg-ssd
|
300×300
|
246ms
|
yolov3
|
608×608
|
582ms
|
Note: the soft core of FZ3 Card is in continual upgrading and its
performance will be improved simultaneously. Different
versions of the same
network have different requirements on computing power. If you have specific
project application, please
contact MYIR for customized optimization.
Video: Demo: AI Fruits Detection Based on MYIR’s FZ3 Card
https://youtu.be/3QoidpG1ERQ

The demo displayed
in the video adopts object detection model ->MobileNet-Yolov3 which is a
deep learning network model with relatively low computational complexity. This
model is suitable for mobile and embedded edge devices with limited computing resource. Data Set adopts MS COCO (Common Objects in Context). This model
performs well on FZ3 deep learning computing card, when inputting images of
416x416 pixels, the average single frame time consuming is 88 ms.
This demo
will be released as open source later, you may download and deploy it to FZ3 Card
and see how it works then.
Introduction of Model MobileNet-Yolov3
Adopts MobileNet as the backbone architecture of the Yolov3 model, Mobilenet-Yolov3
model not only ensures the running efficiency on the equipment with limited
computing resources, but also ensures the accuracy of object detection. The
detection process of the model is shown in the figure below:

Adopts Mobilenet
as the backbone structure to replace Darknet53 of the Yolov3 mode, MobileNet
mainly uses grouping convolution and point convolution to replace the original
standard convolution, can reduce the convolution operation part of the backbone
network greatly, so that the overall computing amount of the network is greatly
reduced. In this model, in addition to turning the backbone architecture into a
more lightweight network MobileNet, other processing procedures are the same
with Yolov3. During which the 11th and 13th Pointwise convolutional layer
output feature maps are extracted respectively, and combined with the final
output feature maps of the backbone for multi-scale prediction.
In
PaddlePaddle framework, Mobilenet-Yolov3 model is further optimized. The
maximum use of tailoring, distillation and other optimization strategies can
make the model compressed by 70% and the reasoning speed increased to two time!
Introduction of Data Set COCO
(Common Objects in Context)
The
COCO (Common Objects in Context) is a large image data set released by
Microsoft. It is designed for object detection, segmentation, human key points
detection, semantic segmentation and subtitle generation. This data set takes
scene understanding as the target and mainly intercepts from complex daily
scenes. Targets in the image are calibrated by precise segmentation. Images include 91 class targets, 328,000 images, and 2,500,000 labels. By far, it has
the largest data set with semantic segmentation, providing 80 categories, more
than 330,000 images, 200,000 of which are annotated, and the number of
individuals in the whole data set is more than 1.5 million.
The COCO
dataset is available at https://cocodataset.org/
About
MYIR
MYIR Tech Limited is a global provider of ARM hardware and software
tools, design solutions for embedded applications. We support our customers in
a wide range of services to accelerate your pace from project to market.
We sell products ranging from board level products such as development
boards, single board computers and CPU modules to help with your evaluation,
prototype, and system integration or creating your own applications. MYIR also
provide our customers charging pile billing control units, charging control
boards and relative solutions inside of China. Our products are used widely in
industrial control, medical devices, consumer electronic, telecommunication
systems, Human Machine Interface (HMI) and more other embedded applications.
MYIR has an experienced team and provides custom services based on many
processors (especially ARM processors) to help customers make your idea a
reality.
More information about MYIR can be found at: www.myirtech.com
|