A naukri.com initiative
Dzone
5d
331
Image Credit: Dzone
How to Port CV/ML Models to NPU for Faster Face Recognition
- Porting computer vision/machine learning models to an NPU was crucial for faster face recognition on edge devices.
- The NPU proved effective for heavy processing on edge devices, enhancing the speed of face recognition operations.
- Face recognition involves a pipeline including face detection, key point identification, alignment, and template extraction.
- Time constraints for detection, face alignment, and template extraction were set to ensure efficient face verification.
- The hardware used for this project was the OK3568-C, which posed challenges due to its limited processing power.
- Inference on the Rockchip NPU required the use of the RKNN-Toolkit2 for model conversion and optimization.
- Inference on the NPU could be done with Float16 or Int8 representations, each with its trade-offs in accuracy and speed.
- Optimizing models to balance speed requirements and accuracy involved a mix of Float16 and Int8 versions in the pipeline.
- Despite some drop in accuracy due to quantization, the Int8-quantized models were effective for access control use cases.
- In conclusion, utilizing NPUs for CV/ML models significantly accelerates inference speed, outweighing minor accuracy trade-offs.
Read Full Article
19 Likes
For uninterrupted reading, download the app