Example

In this project we investigate one of the most fundamental problems in computer vision and artificial intelligence - detection, localisation and recognition of objects in images (2D and 3D) and video. A method should be very robust similar to the human visual system and easily cope with imaging distortions, occlusions, pose changes, class variation etc. We have focused on our long-term research on Gabor features and currently we have a very effective part-based generative model that is easy to train for new classes and easy to use. We further develop our methodology to compete with state-of-the-art. The detection and classification of visual objects in images is one of the main themes of Vision Group. In particular, we are interested in part-based methods that break the problem into two sub-problems: 1) detection of local object parts (eye centre, nostril etc.) and 2) detection of a whole object based on parts' constellation.

Most of our methods use 2D models of objects to detect them from images (2D-to-2D), but we push research towards more natural representations, i.e. understanding 3D variation of objects in 2D images: 3D-to-2D models and methods.

People

Ekaterina Riabchenko E-mail Research Student
Dr. Ke (Cory) Chen E-mail Post-doc Researcher
Joni Kamarainen E-mail Professor

Selected Publications

Learning Generative Models of Object Parts from A Few Positive Examples

By Ekaterina Riabchenko, Joni-Kristian Kamarainen and Ke Chen In 22th Int. Conf. on Pattern Recognition (ICPR2014) 2014.

Our method can be used to learn probabilistic models of annotated object parts using Gabor filters and Gaussian mixtude models.

Download: PDF, BibTex, Source Code

Density-Aware Part-Based Object Detection with Positive Examples

By Ekaterina Riabchenko, Joni-Kristian Kamarainen and Ke Chen In 22th Int. Conf. on Pattern Recognition (ICPR2014) 2014.

Our method learns a probabilistic (generative) model of an object class. The model is part-based where the parts are learned using our part detector method and the full model is learned using a constellation model in canonical space (geometric invariance).

Download: PDF, BibTex, Source Code

Links

Data sets

Caltech-101
"Baseline" set
Pascal VOC
The annual Pascal VOC challenge
LabelMe
Thousands of labelled images in WordNet hierarchy
ImageNet
Millions of labelled images in WordNet hierarchy (the preferred dataset)

Authors and teams

Pedro Felzenszwalb
The Part-Based Method for Object Detection and Classification
Fei-Fei Li
The creator of the Caltech-101 and ImageNet
Andrew Zisserman's group
Seminal works on the Visual Bag of Words (BoW) approach from 2003 (currently rather a "Bag of all tricks we know plus SVM")