Researchers improve method to detect faces in a crowd from small images
Researchers at Carnegie Mellon University have developed a method for detecting multiple faces in a crowd from small images with a greatly reduced error factor, according to a report in The Tartan.
The process described in the Finding Tiny Faces report involves focusing a computer on specific pixels within a “massively-large receptor field,” in which 99 percent of the template examined is beyond the object of interest, and which provides the context to detect small objects.
While most recognition approaches are claimed to be applicable to facial images of various sizes, the study authors say recognizing face images 3 pixels tall is fundamentally different than recognizing those 300 pixels tall. They used foveal descriptors, which are features indicating a specific area to focus on, to recognize small facial images, and separate detectors for different scales and aspect ratios.
The technique produced an AP of 81 percent when applied to the “WIDER FACE” detection benchmark, while existing methods range from 29 to 64 percent, according to study authors Deva Ramanan and Peiyun Hu, a robotics professor and Ph.D. student, respectively.
Ramanan told The Tartan that the detection of faces in a crowd from small images is like “spotting a toothpick in someone’s hand.”