[Imageworld] Post-Doc position on Object and Action Recognition at INRIA Grenoble (LEAR team)

[Remi Ronfard]

The LEAR team at INRIA Grenoble is looking for a qualified post-doctoral 
researcher with a specialization in Computer Vision and Machine 
Learning, on the topic of discovering relationships between actions and 
objects.

The position is offered at the "Rhone-Alpes" Research Unit of INRIA, 
located near Grenoble and Lyon, in France. The Unit includes more than 
600 people, within 26 research teams and 10 support services.

Deadline for applications: June 30, 2009.

Starting date: September 2009

Monthly salary after taxes : 1 983 EUR (medical insurance included)

Contact: Remi.Ronfard at inrialpes.fr

Activities

Recently, a number of image ranking approaches were proposed that build 
upon visual words similarity networks (i.e. [3,4]). These methods 
explore relationships between object categories by analyzing 
similarities of the extracted visual features. In the case of video 
actions, the relationships are more complex as similarities can be 
observed in the spaces of image features, motion features, and also in 
the joint space of image and motion features. An approach to discovering 
relationships in such networks would allow for recognition of objects, 
motions, and human-object interactions. The initial investigation can be 
performed along the lines in [3,4].

In order to achieve the above goal, a good feature extraction method has 
to be developed. Existing spatio-temporal features describe information 
of a video subvolume of a simple shape. Intuitively, the procedure that 
discovers the shapes of such subregions should be guided by some general 
measure of the subregion descriptiveness.

Unfortunately, straightforward extensions of the common 2D subregion 
extraction methods [1] may not be appropriate. Additionally, approaches 
to obtaining good descriptors of the extracted subregions should be 
investigated., with special care taken to obtain good view and 
time-invariant spatio-temporal descriptors.

In order to investigate the relationships between actions and objects, 
the problem of analyzing human-object interactions should be addressed. 
It would be of significant practical benefit to have a method for 
recognizing interactions from an egocentric camera. Ideally, the 
approach would discover atomic interactions from sequences of long-term 
activities. Some of the possible approaches to implement the idea would 
be to consider the interaction models [2].

Skills and Profile

* PhD degree (preferably in Computer Vision or Machine Learning)

* 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 action recognition, video retrieval or 
object recognition is a plus

REFERENCES

[1] A. Oikonomopoulos, I. Patras, and M. Pantic, Human action 
recognition with spatiotemporal salient points, IEEE Transactions on 
Systems, Man, and Cybernetics - Part B: Cybernetics, vol. 36, no. 3, pp. 
710-719, 2006.

[2] Hedvig Kjellstrasom, Javier Romero, David Martinez Mercado and 
Danica Kragic, Simultaneous visual recognition of manipulation actions 
and manipulated objects, in ECCV (2), 2008, pp. 336-349.

[3] Gunhee Kim, C. Faloutsos, and M. Hebert, Unsupervised modeling of 
object categories using link analysis techniques, in CVPR, 2008, pp. 1-8.

[4] Yushi Jing and Shumeet Baluja, Visualrank: Applying pagerank to 
large-scale image search, TPAMI, vol. 30, no. 11, pp. 1877-1890, 2008.

-- 
Remi Ronfard, Researcher, 
INRIA, 655 avenue de l'Europe, 38330 Monbonnot, France
Office 334 76 61 53 03, Cell 336 71 08 88 81