Pipeline for Large-Scale Computer Vision Applications

Store Face Recognition Pipeline

We built cross-day galleries of staff and customers for over 50 stores using multiple matching techniques based on face/body features, ensemble features, and person track trajectories. Our system performs at an industry-leading position (purity: 99.7% merge rate: 99.0%, staff recall: 99.9%, staff precision: 99%).

Store Face Recognition Pipeline

Car Localization Pipeline

We realized 3D vehicle localization through the fusion of keypoint detection results from multiple views of cameras. It can achieve 95% precision and 96% recall even in hard testing sets containing many partially visible cars.

Car Localization Pipeline