This tutorial shows how to quantize an object detection model, using OpenVINO's Post-Training Optimization Tool API. For demonstration purposes, we use a very small dataset of 10 images presenting people at the airport. The images have been resized from the original resolution of 1920x1080 to 960x540. For any real use cases, a representative dataset of about 300 images would have to be applied. The model used is person-detection-retail-0013
The tutorial covers:
- Quantizing the model with POT
- Comparing the mAP metric on FP32 and INT8 models
- Visually comparing results on FP32 and INT8 models with annotated boxes
- Measuring and comparing the performance of the models
If you have not done so already, please follow the Installation Guide to install all required dependencies.