ResearchBib Share Your Research, Maximize Your Social Impacts
Sign for Notice Everyday Sign up >> Login

PhD position on Attribute-based Objects and Events modeling for Fine-Grained Recognition at University of Caen

Country/Region : France

Website : http://unicaen.fr

Description

NERA (the French Aerospatiale Lab) and the Computer Vision group of the University of Caen (France) are opening a PhD position on Attribute-based Objects and Events modeling for Fine-Grained Recognition.
Start date: October 2015.
Location: the student will be located at the Palaiseau Center of the ONERA (nearby Paris) or/and at the University of Caen, France
Contacts: Stéphane Herbin (stephane.herbin-AT-onera.fr) and Frederic Jurie (frederic.jurie-AT-unicaen.fr)
Profile:
* Recent Master (preferably in Computer Science, computer vision, machine learning or Applied Mathematics)
* Solid programming skills; the project involves programming in Matlab and C++
* Solid mathematics knowledge (especially linear algebra and statistics)
* Creative and highly motivated
* Fluent in English, both written and spoken
* Prior knowledge in the areas of computer vision, machine learning or data mining is a plus (ideally a Master thesis in a related field)
Please send applications via email, including:
* a complete CV including a list of publications
* graduation marks as well as rank
* the name and email address of three references
to Stéphane Herbin (stephane.herbin-AT-onera.fr) and Frederic Jurie (frederic.jurie-AT-unicaen.fr) before the end of April 2015. Applicants can be asked to do a short assignment in order to demonstrate their research abilities.
---
Topic: Attribute-based object and event modeling for fine grained recognition from scarce or heterogeneous training data
The objective of the thesis is the design of an automatic system for visual recognition of objects, actions or events, explicitly taking into account the scarcity and heterogeneity of the reference data used to build the models.
The performance of visual object recognition algorithms are increasing each year. Recent achievements rely on efficient statistical learning techniques but also on the availability of massive annotated databases (hundreds of thousands of images). The so-called “deep learning” algorithms in particular have recently shown a significant gain on several object recognition benchmarks.
In real situations, having access to massive and statistically relevant data is a prerequisite that is seldom satisfied: available data are often from disparate contexts, of various image qualities and from heterogeneous viewing conditions. In practice, the fact that representative data is scarce limits the usage of massive learning techniques such as described in the recent literature.
One alternate approach to allow a better consideration of the heterogeneity of training data is to design object models from an intermediate representation of attributes computed on images. These attributes describe object components (wheel, headlight, eye, chin, right arm, waving hand), global configurations (side view, sitting, standing), intensive (color) or extensive (length, surface, duration) features. Models using attributes are generally provided with a probabilistic structure for controlling their detection uncertainty and their correlations or dependencies. If global classes are difficult to estimate statistically from a limited volume of data, it is however possible to learn shared attributes over a larger database.
The use of attributes has two major advantages: 1/ it defines models by description, as opposed to a model from a set of examples, and makes possible the modeling without reference data ("zero-shot learning") 2/ it can make the contextual quality of the model visually understandable through adequate visualization schemes of errors or uncertainties. Attribute visualization leads more directly to interactive approaches in which human cognitive abilities can be recruited naturally.
The aim of the thesis is the design and evaluation of an approach for introducing new models of objects, actions or events in a visual recognition system from a description by attributes and / or a reduced number of reference examples. One of the questions that will be addressed is how to articulate efficiently advanced learning techniques (“deep learning”) and attribute-based models for object recognition, with the possible use of human interaction during the modeling stage.
The work will be an opportunity to address a number of issues of interest to the practical use of recognition algorithms for which the literature does not provide yet good solutions: how to add a new class to be recognized in the system? What are the elements that differentiate this new class from others? How to update the classifiers of the other models to prevent regression? Can we anticipate or guarantee the new model recognition performance? How to incrementally improve performance when new benchmark data become available?
Keywords: visual object and event recognition, visual attributes, fine grained recognition, zero-shot learning, human-computer interaction.
Short bibliography:
? Branson, S.; Van Horn, G.; Wah, C.; Perona, P. & Belongie, S. The Ignorant Led by the Blind: A Hybrid Human?Machine Vision System for Fine-Grained Categorization, International Journal of Computer Vision, Springer US, 2014, 108, 3-29
? Duan, K.; Parikh, D.; Crandall, D. & Grauman, K. Discovering localized attributes for fine-grained recognition Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, 2012, 3474-3481
? Fu, Y.; Hospedales, T. M.; Xiang, T.; Fu, Z. & Gong, S. Transductive multi-view embedding for zero-shot recognition and annotation Computer Vision--ECCV 2014, Springer, 2014, 584-599
? Kumar, N.; Berg, A.; Belhumeur, P. & Nayar, S. Describable Visual Attributes for Face Verification and Image Search Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2011, 33, 1962-1977
? Jayaraman, D. & Grauman, K. Zero Shot Recognition with Unreliable Attributes, NIPS, 2014
? Lampert, C., Nickisch, H., & Harmeling, S. (2013). Attribute-based classification for zero-shot learning of object categories.
? Liu, J.; Kuipers, B. & Savarese, S. Recognizing human actions by attributes Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, 2011, 3337-3344
? Palatucci, M.; Pomerleau, D.; Hinton, G. E. & Mitchell, T. M. Zero-shot learning with semantic output codes Advances in neural information processing systems, 2009, 1410-1418
? Parikh, D. & Grauman, K. Interactively building a discriminative vocabulary of nameable attributes Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, 2011, 1681-1688
? Parkash, A. & Parikh, D. Attributes for classifier feedback. Computer Vision--ECCV 2012, Springer, 2012, 354-368
? Su, Y., & Jurie, F. (2012). Improving image classification using semantic attributes. International journal of computer vision, 100(1), 59-77.
? Vedaldi, A.; Mahendran, S.; Tsogkas, S.; Maji, S.; Girshick, R. B.; Kannala, J.; Rahtu, E.; Kokkinos, I.; Blaschko, M. B.; Weiss, D. & others Understanding Objects in Detail with Fine-grained Attributes IEEE Conference on Computer Vision and Pattern Recognition, 2014
? Yu, F. X.; Cao, L.; Feris, R. S.; Smith, J. R. & Chang, S.-F. Designing category-level attributes for discriminative visual recognition Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on, 2013, 771-778

Last modified: 2015-03-26 22:38:22