This article demonstrates how to build a computer vision solution for detecting car positions in images using two approaches: Amazon Rekognition and a custom SageMaker model.

• Amazon Rekognition detects car wheels, then infers orientation using rule-based postprocessing
• Custom Detectron2 model trained on SageMaker for more granular car part detection
• Solution uses Lambda functions behind API Gateway for serverless deployment
• Detectron2 fine-tuned on car parts segmentation dataset with ~500 annotated images
• Web application built with AWS Amplify allows users to upload images and select detection method
• Trained model hosted on SageMaker real-time endpoint using GPU instance
• Rekognition Custom Labels option available for improved wheel detection accuracy
• Complete infrastructure provisioned using AWS CDK for easy deployment

The article provides a flexible, serverless architecture for car pose detection that can be adapted for other object orientation use cases, with choice between managed service or custom ML model approaches.