Our technology uses a deep learning neural network to determine landmarks for every face found in a picture. The model is pre-trained and optimized to run natively on the mobile and PC (with no internet required), thus preserving the privacy of users without uploading the photos to the cloud. It can scan a large number of photos (in thousands) and group all the detected and recognized faces into unique faces (unique persons). Adult faces, in particular, are detected with a high degree of accuracy as facial landmarks for adults will be distinct. The model works well in the presence of partial occlusions that cover a face. This technology comes as an SDK with an extensive list of APIs to cater to 3 distinct use cases. SDK is a plug and play that can be quickly integrated into a mobile/PC environment.
This technology has three distinct use cases, which are described below.
FACE RECOGNITION: Recognize a person with different facial expressions, partial occlusions, different lighting conditions and with facial hair, headdress, and sometimes even with eyeglasses.
FACE CLUSTERING: All the faces are grouped using a hierarchical clustering algorithm. This results in classifying the entire gallery in a phone, or a folder containing thousands of pictures into buckets of unique faces. Threshold and minimum group size provide flexibility to tune the algorithm for finer control. The second level of clustering allows relationship tagging by mapping the frequency of occurrences with another person.
FACE SEARCH: Find a suspect in a mass-participation venue with a live video stream, where the face recognition engine running on the mobile prompts a security person in real-time when a suspect is detected. On the other vertical, use a single face shot to get all the photos of a person from a large data set such as a marathon, school event or theme park.
Many photo printing applications rely on detecting faces and grouping faces of the same person together. The SDK comes handy to group photos belonging to the same person. For example, a school’s repository of photos over a year can be quickly indexed for each pupil. A photography company covering a marathon with multiple photographers can separate the photos belonging to the same person at the end of the event, where these photos can be further customized and used in several print applications. Similarly, photographers in a zoo or amusement parks need not issue physical tickets to the users as the face recognition system will be able to group all the photos belonging to a person/family into a single folder, which the user can check out at the exit gate before leaving the venue.
The same face recognition model can power security applications to detect unassuming suspects in malls and fairs where a large number of people congregate. A list of suspect (person of interest) photos can be compared with a live streaming body-camera/CCTV to prompt security personnel to interrupt the suspect when the face recognition system positively identifies a person of interest.
There are several off the shelf face recognition models in the market, but most of them rely on cloud computing and requires uploading photos to the cloud. Our model works in a closed environment natively on a PC or Android/iOS environment with zero data connectivity and with a high degree of efficiency. For applications/systems, where privacy is a major concern, this solution comes handy.
Many Face Recognition Models are computationally demanding and modern day mobiles still struggle to cope up with real-time performance. The company have designed a lite model with a high degree of accuracy to recognize and index all photos of the same person.
This model comes as an SDK that can be quickly deployed in both iOS and Android platforms to power many applications that rely on detecting and recognizing faces.
The company have a slew of proof of concept applications to showcase the technology.