NEC's Software Detects Human Faces in Video, Sorts Them Out

Jan 15, 2008
Yousuke Ogasawara, Nikkei Electronics
NEC's software that detects people appearing in a video and displays lists of their faces
NEC's software that detects people appearing in a video and displays lists of their faces
[ If it clicks, the expanded picture will open ]

NEC Corp has developed a software technology to detect main people appearing in a video and display lists of their faces. It helps users easily search videos they want to see from video archives, NEC said.

The technology can be applied to home appliances, mobile devices, video sharing Websites and video archives for companies, etc, according to NEC. The company said it will advance development with a view to commercializing the technology in the near term.

After grouping the faces of people derived from video by person, the software chooses a representative, front face view from each group. Then it automatically forms and displays lists of faces aligned in the order of appearance frequencies by video file.

Seeing them, users can find out who will show in each video. In addition, users could recognize what faces are contained in the video even if they did not know their names, NEC said.

To group detected faces by person, the software implements clustering (categorization by the degree of similarity) in two phases. First, the software groups each person's face in different angles, expressions and lighting conditions when the video was recorded, for example, deriving them from one shot, which starts when the camcorder was turned on and ends when it is turned off.

Then, the software chooses every front face view from several shots and compares them across groups. If similar faces are found in different groups, they will be included in one group and recognized as faces of an identical person.

Through this measure, the software can detect faces generating less overlap and failure, NEC said. If faces of five people are displayed in a video list, the software reportedly succeeds in detecting people in a video at about 80%.