Visual Place Categorization

I am in the organizing committee of the VPC'09 workshop associated with CVPR 09. For introduction to VPC, see the workshop page. We are providing a new VPC dataset for this new problem, which can be downloaded from the workshop dataset page.

VPC refers to identify the semantic category of a place using visual information collected by an autonomous robot platform.

Our paper on VPC is accepted by IROS 2009 (paper [C9]).

libHIK : clustering and classifying histogram feature vectors

We present a clustering method for feature vectors that are histograms, utilizing the Histogram Intersection Kernel (HIK).

We also provide functions that make the testing of SVM models on histograms extremely fast.

We apply these methods in bag of visual words model to generation visual codebooks for histogram feature vectors. Our method consistently improves recognition accuracy in object recognition, scene recognition, sports event recognition, and visual place categorization (VPC). This method is presented in ICCV 2009 (paper [C8]]).

More details here. Or, download libHIK directly here. Please cite our ICCV 2009 paper ([C8] in my publication page) if you use libHIK.

libHIK v 1.0 was published July 30 2009, updated to v 1.5 on August 10, 2009.

CENTRIST: A Visual Descriptor for Scene and Place Recognition

We present a visual descriptor called CENTRIST (CENsus TRansform hISTogram) that is suitable for recognizing the semantic category of natural scenes and indoor environments, e.g. forests, coasts, streets, bedrooms, living rooms, etc.

Recently we focused on understanding why CENTRIST is suitable for place and scene recognition. Our results are shown in paper [W3].

CENTRIST was used used with PCA for scene recognition. More details here. C++ source code is available. Please refer to paper [C7].

CENTRIST was also used within the bag of visual words (BOV) framework. CENTRIST also shows superior performance for scene recognition. Please refer to paper [C8].

Synergistic Activity and Object Recognition

We use RFID and Vision as two complementary sensors in a Bayesian network, and recognize daily activities & objects simultaneously.

Since RFID signals are used as noisy labels, absolutely no manual labeling of images is need (neither for activity segmentations nor for object labels).

More details here.

Fast and Robust Face Detection / Rare Event Detection

We propose FFS to replace the AdaBoost classifier in the Viola-Jones cascade face detector. FFS is about 2 orders of magnitude faster than their original AdaBoost implementation.

The classifiers in a cascade has different learning goals than usual classifiers. We formalize the asymmetric learning goal in a cascade and provides inference algorithm for it.

More details here, C++ source code and a demo program/video included.

Effective and Efficient Sampling Strategies for Imbalanced Learning problem

We propose two methods EasyEnsemble and BalanceCascade to deal with learning problems with imbalanced datasets.

These methods are very efficient – their training time is almost identical to that of the down sampling method (which only use a small portion of the original data set, thus is very fast)

They are also effective – they pick up the information that's thrown away in down sampling by using multiple samples.

See the publication page for more details (paper [J6] and [C5]).

Selective Ensemble of learners – many could be better than all

We show that in a ensemble classifier (e.g. bagging and boosting), it is not always the best to use all the weak classifier in the ensemble. We show that by choosing only a subset of the weak classifiers, the new ensemble will have better accuracy.

See the publication page for more details (paper [J2], [C1] and [C2]).