Efficient Lightweight Multi-Source Domain Adaptation for Person Re-ID via Self-Paced Meta Learning

What is it about?

Cameras widely used in public places need technology to accurately identify the same person across different monitoring screens, which is known as person re-identification. Traditional recognition tools either require massive manual labeling work or struggle to work well when camera angles, lighting and scenes change, and most existing models are too bulky to run on small edge devices like regular surveillance cameras. To fix these problems, we created a lightweight AI model for cross-camera person recognition. Our model learns useful rules from multiple sets of monitoring data step by step, and combines overall and local details of pedestrian images to boost recognition accuracy. Test results on both public standard datasets and real-world railway station monitoring data prove that our solution runs fast with low power consumption, and maintains excellent recognition performance in complex real scenarios. This work helps make intelligent monitoring technology easier to deploy and use in daily public security systems.

Featured Image

Photo by BENJAMIN CHAN on Unsplash

Photo by BENJAMIN CHAN on Unsplash