Computer Vision ML Intern Project

Question

Hello everyone.

I am an intern beginning a new project focusing on computer vision through a Machine Learning Model. Our goal is to detect based on images whether the object in focus is properly aligned or not. For example, if the object is a cube, we want to be able to classify a certain angle that the cube is sitting at as a pass with a specific tolerance, and that every other angle is a fail. We also hope to output the difference from the pass tolerance.

We looked into available APIs already built in to the AWS Rekognition platform, but found none that seem to be able to do object positional analysis. Any sort of guidance on how we could approach this project with ML and the tools that at our at disposal that would make developing this model as efficient as possible will be much appreciated!

Thank you!

AWS TAM · Accepted Answer

Hi Shimekan,

Here is how I would approach object positional analysis using AWS Rekognition:

Amazon Rekognition's object and scene detection capabilities can identify the location of common objects in images and videos by returning bounding box coordinates. This can be done using the DetectLabels API.

The bounding box information returned by DetectLabels can be used to infer the position and orientation of the detected objects. For example, you could analyze the relative positions and sizes of the bounding boxes to determine if an object is properly aligned.

To implement this, you would need to:
        Use the
        DetectLabels
        API to detect objects in your images
        Extract the bounding box coordinates for the objects of interest
        Analyze the bounding box data to determine the object's alignment and orientation
        Compare the observed orientation to your desired "pass" tolerance to classify the object as aligned or not

While Amazon Rekognition does not provide a pre-built API for this type of object positional analysis, you can build a custom solution using the existing Rekognition capabilities along with other AWS services like AWS Lambda, Amazon S3, and Amazon SageMaker.

I recommend reviewing the Amazon Rekognition documentation for more details on the DetectLabels API response structure and how to work with the bounding box data. You can find the latest documentation at the [Amazon Rekognition Developer Guide](https://docs.aws.amazon.com/rekognition/latest/dg/what-is.html)

Mi_Sha · Answer

You can also train your own model in Amazon SageMaker. You can use build in models or bring your own model to fine-tune it on your images

![Enter image description here](/media/postImages/original/IMX1XGJnWxTRKtVlg0NKKQgQ)

Computer Vision ML Intern Project

Relevant content