In this project we address the problem of recognizing human actions in a video sequence. Unlike previous approaches, we aim at developing a method able to continuously recognize and segment actions. For this purpose a per-frame rather than a per-video representation is needed. This means that the data (short videos) are represented as time-series of vector-valued observations.
Fisher kernels have been proved quite powerful when used with Gaussian mixture models (GMM) in classification/recognition problems. However, when structured objects such as time-series have to be classified, the independence assumption that is typically enabled with a GMM violates the temporal coherence of data. The drawback is that rich time-dependent information is lost and used neither for training nor for recognition. In this project, the use of Fisher kernels in conjunction with a generative model that encodes the temporal nature of data (e.g., Hidden Markov Model or HMM) will be initially investigated in the context of the human action/gesture recognition from color and/or depth image sequences. Further, such a model will explain the generation of action/gesture primitives of a dictionary that can compose any complete action/gesture of a predefined action set.
The project is suitable to a second year master student with very good background in statistics and probability theory, machine learning and computer vision, as well as very good programming skills in Matlab. The project is part of a broader effort to build computational models for human-robot interaction and is funded by the ERC advanced grant VHIA.
The project may start anytime after 1 February 2015 for a period of six months and it may continue with a Phd thesis, which is fully funded by the VHIA grant.
Information for applicants: Please send your complete CV, university grades, and the names and emails of two recommending persons to firstname.lastname@example.org. Students enrolled in a French university will receive a monthly net salary of 430€. Students enrolled in another university will receive a monthly net salary of 1100€. Please note that some restrictions apply to non-French students and their admission is conditioned by an approval from the French Ministry of Defense.
 T. Jaakkola, D. Haussler: Exploiting generative models in discriminative classifiers, NIPS, 1999.
 F Perronnin, C Dance: Fisher kernels on visual vocabularies for image categorization, CVPR 2007
 L. van der Maaten: Learning Discriminative Fisher Kernels, ICML 2011
 C. M. Bishop. Pattern recognition and machine Learning, Springer-Verlag, 2006