Topic Title: Face and Person Recognition across Large Pose
Identity recognition is a key component in many intelligent applications, such as video surveillance, business analytics, financial security, access control and so on. Especially for business analytics in physical store, we could track the consumers’ preferences based on their identity. Identity and big data could be combined further to enhance consumers’ shopping experiences. Currently, identity could be obtained through mobile APPs, Bluetooth, Wi-Fi or video camera. Among them, face (or person) recognition based on video camera is the most user-friend and device-free technology.
For user-friend experiences and easy deployment, we want to develop high performance face/person recognition algorithm, especially in crowded store or supermarket. To avoid person occlusion in crowded environment, video cameras are usually install at height > 3m in very large pitch angle (> 30 degree). In practice, we find that the large angle installment significantly degrades the recognition rate. Although many pose robust face recognition algorithms were proposed in the existing works, they mainly focus on the yaw angle. Dealing with both pitch and yaw angles together is still a challenge.
As compared with face recognition, person recognition (or re-identification, re-id) is naturally more robust to the pose variations. This is because person recognition is mainly based on holistic features, e.g., color and texture of clothes. Therefore, combing both face and person features to do recognition could be a reasonable research direction in order to improve the recognition rate across large pose.
We are looking for the researchers who are experts either in this field or the related fields and also keen to aim for the challenges and opportunities on this topic.
Aiming to “New Retail” scenarios, advanced large pose face and person recognition tries to solve the following practical problems:
- Collect a large-scale video dataset, which contains thousands of ID (Identity or subject) and millions of video frames. The bounding boxes and IDs of face and person are labeled in each frame. The dataset should have large quality variations, especially in pitch and yaw angles.
- Develop pose robust face and person detection algorithms.
- Develop efficient long-term face and person tracking algorithms.
- Develop quality assessment and frame fusion algorithms based on the sequence information to improve recognition rate.
- Develop 2D/3D geometric methods to normalize the pose of face and person image.
- Develop robust and discriminative metric learning methods for face and person matching.
- Build up a fast and accurate identity recognition system in crowed environments by combing both face and person biometrics.
Related Research Topics
Many sub-problems related to this topic have been studied for many years in academy and industry, which are listed as follows:
- Object detection
- Multiple object tracking
- Image quality assessment
- Face recognition
- Person Re-Identification
- Set based face/person recognition l Multi-modal biometrics fusion
- Face 3D morphable model (3DMM)
- Skinned multi-person linear model (SMPL)