Object Detection Bounding Box gone wrong

[arafattehsin]

System information

• Windows 10:
• .NET Core 3.1:

Issue

• I created an object detection model with fantastic accuracy
• All works good. However, when I try to map the bounding box coordinates, I see that the coordinates returned are not correct by the model (or may be I am totally wrong and I should know how to make it work)
• The bounding box should have been drawn correctly around the detected items.

Source code / logs

This is the Preview I get (before training)

This is what I get after training.

This is what I get for a different image (after training).

Note: I know they are delicious. 🍩

However, after passing the same image to model, I get the coordinates which are mapped like this;

Nothing fancy in terms of code;

 var boundingBoxes = predictionResult.BoundingBoxes;

    foreach (BoundingBox box in boundingBoxes)
    {
      // read image file
      Image oldImg = Image.FromFile(destinationPath);
    
      using (Graphics g = Graphics.FromImage(oldImg))
      {
          // Create pen.
          Pen redPen = new Pen(Color.Red, 3);
    
          // Draw Rectangle
          Rectangle rectangle = Rectangle.FromLTRB(Convert.ToInt32(box.Left), Convert.ToInt32(box.Top), Convert.ToInt32(box.Right), Convert.ToInt32(box.Bottom));
    
          g.DrawRectangle(redPen, rectangle);
      }
    .
    .
    .
    .
    
    }

Please tell me that I am absolutely wrong by correcting me or posting a correct solution. Thanks! 🙏

[JakeRadMSFT]

Hello!
It looks like you're using Model Builder. Would you be up for trying out our preview build? I believe this issue is resolved there and a public release is coming soon.

[arafattehsin]

Hey @JakeRadMSFT thanks for your response. Do we have an update on the Model Builder lately? I can try in an hour and get back to you then.

[arafattehsin]

Hey @JakeRadMSFT - I just checked that I have the latest version:

It matches with the latest atom.xml;

May I know if I am doing anything wrong?

[LittleLittleCloud]

Hi @arafattehsin

The coordinate values in BoudingBox are normalized to 800L * 600H dimension. This is because in the training pipeline all images are resized to 800 * 600. So you will need to do a re-mapping for BoundingBox coordinate values so they can match the custom image size. Specifically, for BoundingBox.Left and BoundingBox.Right, you need to divide it by 800 and multiply the width of the custom image. And for BoundingBox.Top and BoundingBox.Bottom, divide it by 600 and multiply the height of your image.

[LittleLittleCloud]

@JakeRadMSFT @briacht
Looks like the normalized bounding boxes are mentioned nowhere so maybe we should add a comment to output class?

[arafattehsin]

Hey @LittleLittleCloud - you're the star. It works fine. Thanks a lot!