A neural network finds faces in photographs
In November 1995 Henry Rowley, Shumeet Baluja and Takeo Kanade of Carnegie Mellon described a face detector in a technical report: a neural network examines every 20 × 20 pixel window at several scales, and arbitration between two networks weeds out false detections. On 130 complex images with 507 faces the main system found 85.4% of them, with one false detection per 1,319,035 windows.
Why it matters
Finding a face in an arbitrary photograph, not an identity photo, became a task with a shared test: the authors gave their 130 images to other researchers as a standard for comparison. As the "MIT+CMU test set", it is where Viola and Jones would show their cascade in 2001, comparing it with this very detector.
Technical report CMU-CS-95-158R, 26 pages. The network was trained on nearly 1050 aligned faces from the CMU and Harvard databases, each turned into fifteen variants by rotation, scaling, shifting and mirroring; non-faces were gathered by bootstrapping, the current network’s false detections being fed back into training. The test combined three sets: 42 CMU images with 169 faces, Sung and Poggio’s 23 images with 155, and another 65 images with 183, together 83,099,211 windows. From 78.9% of the faces with eight false detections to 92.7% with 1768 for a single network; the system with 362 found 90.5%. Two networks took about 590 seconds for a 320 × 240 image on a Sparc 20, the fast version about 24; a demonstration ran on the Web. The journal version, IEEE PAMI 20(1), January 1998, pp. 23-38, gives 77.9% to 90.3% on the same images and the main system at 86.2% with one false detection per 3,613,009 windows, adds a test on a FERET subset (97.8% to 100%) and 383 seconds on an SGI Indigo 2. Its introduction still says 90.5%, but that is the 1995 report’s figure: the journal’s table gives at most 90.3%.