• State
  • Locked Locked
  • Replies 1 reply
  • Subscribers 99 subscribers
  • Views 74 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

SK-AM62A-LP: object detection not working well

Adithyan DV
Prodigy 30 points
Part Number: SK-AM62A-LP

Tool/software:

Hi,

I've successfully used Edge AI Model Maker to develop an object detection model, but I'm not getting the expected outcome. The model name appears in the apps_python terminal print ncurse, but I'm not receiving the bounding box or a recognised class in my output.

Attaching my config file

 

title: "Demo"
                                                          
inputs:                                                  
  input0:      
    source: /dev/video-imx219-cam0     # IMX219 raw sensor
    format: rggb                        # raw Bayer input
    width: 1920                                          
    height: 1080                                          
    framerate: 30                                                 
    subdev-id: /dev/v4l-imx219-subdev0  # Required by ISP
 
                                           
models:               
  model0:
    model_path: /opt/model_zoo/SAMPLE-OD-CHECK
    viz_threshold: 0.6
outputs:       
  output0:      
    sink: kmssink
    width: 1920
    height: 1080     

flows:         
  flow0: [input0,model0,output0]

regards 

Adithyan DV

  • 0
    TI__Genius 15086 points

    Hi Adithyan,

    I understand your model is not performing on live data as accurately as you expect. I assume the imaging pipeline itself is running fine with IMX219 CSI camera, and that you are indeed seeing a live feed on your attached monitor. (Note that detection information does not print in the logs by default)

    Firstly, let's consider how your model did on the original data during modelmaker. Within edgeai-modelmaker, you will find that you model's final version is hosted under a path like:

    • edgeai-modelmaker/data/projects/$PROJECT_NAME/run/$DATE_STRING/$MODEL_NAME/compilation/$SOC/work/$MODEL_SHORTNAME
    • There will be an 'outputs' folder that holds the postprocessed version of some of your dataset. Do the images here include the expected detections?
    • Alternatively, consider what the final training accuracy was for your model. These accuracy values would be in a training "run.log" file

    Assuming the above is correct, I would suggest changing the input source for your model to point to a directory of image files for your dataset. Please verify it is responding well to familiar data. 

    Best Regards,

    Reese