We present a novel boosting cascade based face detection framework using SURF features. The framework is derived from the well-known Viola-Jones (VJ) framework but distinguished by two key contributions. First, the proposed framework deals with only several hundreds of multi-dimensional local SURF patches instead of hundreds of thousands of single dimensional haar features in the VJ framework. Second, it takes AUC as a single criterion for the convergence test of each cascade stage rather than the two conflicting criteria (false-positive-rate and detection-rate) in the VJ framework. These modifications yield much faster training convergence and much fewer stages in the final cascade. We made experiments on training face detector from large scale database. Results shows that the proposed method is able to train face detectors within one hour through scanning billions of negative samples on current personal computers. Furthermore, the built detector is comparable to the state-of-the-art algorithm not only on the accuracy but also on the processing speed.

[Dataset download]

1) The file faceneglist.7z contains the image file list that this paper used for negative training set (a small portion of in-house collections are removed due to privacy issues).

2) File [facepos.7z] contains the cropped images that this paper used for positive training set (a small portion of in-house collections are removed due to privacy issues).

The readme file "facepos_readme.7z" below gives a detailed readme and image-list on where the positive face images coming from.

Note the facepos_readme.7z is an extended dataset used for the tech-report, and contains more faces than facepos.7z. See readme inside for details.

3) Please read the readme file in each package before using it.

4) If you use these two collections for any research purpose, please cite our paper.

An extended version is available HERE.

An FAQ on how to implement it is also available on the same page.