Home

Course

Research

Publication

Honors

Misc

1. Fast Asymmetric Learning for Cascade Face Detection

     A cascade face detector uses a sequence of node classifiers to distinguish faces from non-faces. The main advantage is its testing speed: a cascade detector can detect faces in real time. Cascade detectors can be applied to detect objects other than faces.

     We improved the cascade detector in the following aspects:

           1) Faster Training algorithm: We proposed Forward Feature Selection (FFS) to train node classifiers in a cascade. The original cascade face detector used AdaBoost, and took a few weeks to train, while FFS used about 3 hours. The cascade detector trained by FFS had approximately the same performace (measured by detection rate/false positive rate) as the cascade trained by AdaBoost. We also presented a faster implementation of the AdaBoost algorithm in training a cascade face detector, which took about 8 hours.

           2) Asymmetric inference algorithm: Node classifiers has asymmetric learning goals. A node classifier requires very high detection rate (e.g. 99.9%) and only moderate false positive rate (e.g. 50%). We proposed Linear Asymmetric Classifier (LAC) to explicitly handle this asymmetry. Compared with algorithms that do not consider asymmetry, LAC reduced about 20% missed faces in each node classifier, when false positive rate of these nodes are set to 50%.

      More details on our methods can be found in papers (PAMI 2008, IJCV 2008, ICML 2005, NIPS 16). Source code is also available here. See here for a live demo executables under Windows, and a demo video. More descriptions are here.

      More details on our methods can be found in the paper (ICCV 2007).

  3. Place category recognition

     We propose spatial PACT (Principal component Analysis of Census Transform histograms) as the representation to recognize the semantic category of an image. PACT captures both local and global shape of an image.

     More details of PACT can be found in the paper (CVPR 2008) . Source coe is also available here.