• Author Author: ralphjy
  • Date Created: Date Created
  • Views 653 views
  • Likes 2 likes
  • Comments 3 comments
Related
Recommended

Avnet UltraZed-EV Starter Kit Road Test - Vitis AI

ralphjy

Mario Bergeron recently did a tutorial project on Hackster.io vitis-ai-1-1-flow-for-avnet-vitis-platforms  that "provides detailed instructions for targeting the DNNDK samples from the Xilinx Vitis-AI 1.1 flow for Avnet Vitis 2019.2 platforms."   The UltraZed-EV is one of the supported platforms.  I need an AI component for my roadtest project so this was a very timely tutorial.  I decided that since he provided pre-built images that I would do a quick run through of the examples on my board.  I'll demo the examples that I think have the most relevance to my roadtest project.

 

Installation of Vitis-AI embedded package

Installation was straightforward.  The sdcard folder is mounted on /run/media/mmcblk1p1.

  • cd /run/media/mmcblk1p1
  • cp dpu.xclbin /usr/lib/.
  • cd runtime/vitis-ai_v1.1_dnndk
  • source ./install.sh

 

The installation fails to install python support but apparently that was expected.  It's not required for the tutorial but that means the python examples will not work.

The installation process logs you out so a re-login is required.

 

Use the dexplorer utility to verify the installation:

image

 

List of examples:

image

 

Pose Detection

 

Face Detection

 

ADAS Detection

 

Video Analysis

 

The other non-python examples are various different classifiers (inception_v1, mobilenet, resnet50).  There are two examples for each classifier.  The first displays the image it is classifying and outputs the result on the serial console.  This runs very quickly over 500 images so you really can't visually correlate the result but it gives you a feel for the DPU execution time.  The second example runs over 1000 images without displaying and records the overall execution time for classification which gives you an effective average frame rate.

 

Here are the results for mobilenet:

First example - dpu execution time is around 4.5ms per image

image

Second example - overall execution time for 1000 images is about 5.4s for an average frame rate of around 183fps.

image

 

The DPU performance is quite impressive.  The Video Analysis example is the type of AI processing that I want for my project - to detect and classify objects in a video stream.  Now I need to look at Vitis DNNDK development flow and think about how to integrate this on multiple streams with VCU.

 

This was a nice respite from working on PetaLinux build problems image.

 

Links to previous posts for this roadtest:

  1. Avnet UltraZed-EV Starter Kit Road Test- the adventure begins.....
  2. Avnet UltraZed-EV Starter Kit Road Test - VCU TRD
  3. Avnet UltraZed-EV Starter Kit Road Test - VCU TRD continued
  4. Avnet UltraZed-EV Starter Kit Road Test - Port PYNQv2.5
  • in reply to ralphjy

    Ralph,

    Thank you for the response.

    That sounds like the sort of experimentation I should be doing.
    I have a lot of front facing windows on my house, including a window beside the front door.
    Regarding to the UZ7EV-EVCC and VCU functionality,
    you'll be happy to hear that we've made progress on those designs.
    Keep an eye out for updates from

    Regards,

    Mario.

  • in reply to albertabeef

    Hi Mario,

     

    It seems like this was ages ago, but I did try a number of object classification models.  I was primarily interested in identifying what was passing in front of my house.  In general, I could identify humans vs animals and cars vs trucks and it was reasonable at detecting faces - but it was highly dependent on ambient lighting and object orientation relative to the camera.  I basically, concluded that I needed a whole lot more training data and probably augmentation from other sensors (sort of like being a human as you said image).  I had real problems with summer afternoons when there were distinct shadows moving through the image. 

     

    I mentioned that I'm temporarily on a different track, but I'd like to get back to it in a couple of months and I'm interested in seeing if I can apply some of your learning.  I definitely want to augment what I use for detection and classification.  I was thinking about adding audio sensors and location sensors (PIR, LIDAR) so I can identify object position and movement in the image.  I've already used these types of sensors for detection alerts.  Maybe add a second camera to provide classification from a different viewpoint.

     

    I have a couple of hurdles to overcome when I restart, I just recently messed up VirtualBox (so my VMs won't start) and I still have an issue using the VCU with my IP cams (software decoding works).  Hope that fixing VirtualBox won't be too hard, or it will be a painful restart.

     

    -Ralph

  • Great overview Ralph !

    Thanks for the face detection video 

    I am very impressed by Xilinx’s pre-built face detection model.
    Have you tried any of the other models on
    your custom videos ?

    I have tried some models on rock climbing videos ...

    Object classicification videos think a climber is an airplane, and pose estimation models don’t even see the climber ...

    If we take a step back, this is understandable ... most (if not all) of the humans in the training datasets are standing on the ground, not hanging from the walls

    This is a testiment to our magnificent brains, which can generalize objects in the most unusual and unexpected places.

    Would love to get your feedback on this