TensorRT detect.py inference fix

[glenn-jocher]

May resolve #9578

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced testing procedures and visualization improvements for various models in Ultralytics' YOLOv5 repository.

📊 Key Changes

• Extended Continuous Integration Tests: Added tests for exporting .onnx models and subsequent predictions for classification, detection, and segmentation tasks.
• Batch Size Standardization: Unified batch size initialization to 1 for all tasks and set dynamic batch sizes when using webcam data sources.
• Visualization Adjustment: Commented out the drawing of segment polygons during segmentation prediction to possibly remove or update visualization code.

🎯 Purpose & Impact

• Improved Reliability: Testing models in .onnx format ensures compatibility, enhancing trust in model exports for deployment.
• Flexible Input Sources: Adjusting batch sizes dynamically caters to variable input sources like webcams, allowing for more robust and versatile usage.
• Streamlined Visualization: Temporarily disabling polygon drawing could indicate a move to simplify or refactor visualization for better clarity or performance.

These changes can lead to more stable software and a better experience for developers and end-users working with different model formats and input sources. 🚀