Overview / Usage

Project Description: In this project, we conducted a head-to-head comparison between two powerful object detection approaches, YOLOv7 native and Intel OneAPI Libraries, in the context of pothole detection. 🕳️

📌 Objective: Our goal was to determine which approach offered the best trade-off between processing power and accuracy for real-time pothole detection.

🚀 Highlights: Implemented YOLOv7 for pothole detection using PyTorch. Leveraged Intel OneAPI Libraries, achieving a surprising twofold increase in processing power. Conducted comprehensive validation and model conversion steps. Post-quantization with NNCF for enhanced efficiency.

Methodology / Approach

• Get Pytorch model
• Prerequisites
• Check model inference
• Export to ONNX
• Convert ONNX Model to OpenVINO Intermediate Representation (IR)
• Verify model inference
  • Preprocessing
  • Postprocessing
  • Select inference device
• Verify model accuracy
  • Download dataset
  • Create dataloader
  • Define validation function
• Optimize model using NNCF Post-training Quantization API
• Validate Quantized model inference
• Validate quantized model accuracy
• Compare Performance of the Original and Quantized Models

Technologies Used

In the context of YOLOv7 or other object detection models, INT8 quantization can significantly improve inference speed and reduce model size without sacrificing much accuracy, making it a practical choice for many real-time applications.

Repository

https://github.com/Ragulprince/openvino-yolov7