Pattern recognition; Large feature space; Part-based models; Object detection
Dubout C., Fleuret F. (2014), Adaptive Sampling for Large Scale Boosting, in Journal of Machine Learning Research (JMLR)
, 15, 1431-1453.
Dubout C., Fleuret F. (2013), Accelerated Training of Linear Object Detectors, in Proceedings of the IEEE international conference on Computer Vision and Pattern Recognition Workshop
, -, -.
Dubout C., Fleuret F. (2013), Deformable Part Models with Individual Part Scaling, in Proceedings of the British Machine Vision Conference (BMVC)
, -, -.
This project is a one-year renewal of the three-year VELASH project, which started in September 2009. The objective of VELASH is to design novel algorithms to facilitate the development and the use of very large families of image feature extractors for object detection in natural scenes.In coordination with the European MASH project, we have created over the last two years a collaborative platform to allow multiple contributors to develop and test such image feature extractors collaboratively.In VELASH, we have adapted Boosting to deal with multiple families of feature extractors, and have developed new generic procedures extremely efficient in that context. This new techniques have been validated on object classification and detection on canonical databases (INRIA pedestrians, CIFAR, Caltech 101, etc.)The objective of this renewal is to move from monolithic models, and to extend state-of-the-art part-based object detection methods. Our objective is twofold. First, while current methods mainly exploit variants of histograms of gradients -- which result in edge-based templates stable to local deformation -- we want to leverage the multiple modalities provided by the many families of feature extractors we have already implemented. Second, we will investigate the modeling of the joint behavior of features computed on individual parts, instead of relying on strong assumptions of conditional independence.