Face to Face with Efficiency: Real-Time Face Recognition Pipelines on Embedded Devices

While real-time face recognition has become increasingly popular, its use in decentralized systems and on embedded hardware presents numerous challenges. One challenge is the trade-off between accuracy and inference-time on constrained hardware resources. While achieving higher accuracy is desirable, it comes at the cost of longer inference-time. We first conduct a comparative study on the effect of using different face recognition distance functions and introduce a novel inference-time/accuracy plot to facilitate the comparison of different face recognition models. Every application must strike a balance between inference-time and accuracy, depending on its focus. To achieve optimal performance across the spectrum, we propose a combination of multiple models with distinct characteristics. This allows the system to address the weaknesses of individual models and to optimize performance based on the specific needs of the application. We demonstrate the practicality of our proposed approach by utilizing two face detection models positioned at either end of the inference-time/accuracy spectrum to develop a multimodel face recognition pipeline. By integrating these models on an embedded device, we are able to achieve superior overall accuracy, reliability, and speed; improving the trade-off between inference-time and accuracy by striking an optimal balance between the performance of the two models, with the more accurate model being utilized when necessary and the faster model being employed for generating fast proposals. The proposed pipeline can be used as a guideline for developing real-time face recognition systems on embedded devices.