Section: Overall Objectives

Highlights

• Action recognition. LEAR has developed several successful methods for action recognition [7] , [11] , [18] . Our approach for action recognition in still images automatically determines objects relevant for an action given a set of training images [7] . In the PASCAL visual object classes challenge 2011 it achieved best results on three out of ten action classes and the best result on average over all classes.

  The approaches [11] , [18] model the dynamics of actions in videos. In [18] dense trajectory descriptors are extracted and shown to outperform existing video descriptors. In [11] an “actom sequence model” is introduced, which decomposes actions into sequences of (overlapping) action-units called “actoms”. Each actom gathers temporally localized discriminative visual features of the action. This actom sequence model outperformed state-of-the-art approaches on the “Coffee and cigarettes” dataset.

• Large-scale classification. LEAR has designed an efficient and scalable approach for large-scale image classification. The approach [10] allows to gracefully scale up to large number of categories and examples while learning the underlying taxonomy of the categories at the same time, by using a trace-norm regularization penalty. Promising experimental results on subsets of the ImageNet dataset were obtained, where our method outperforms state-of-the-art approaches using 16-Gaussian Fisher vectors. A spatial extension of Fisher vectors [15] allows dimensionality reduction, as does the compression technique presented in [5] .

• INRIA Visual Recognition and Machine Learning Summer School. This year we co-organized the second edition of the ENS-INRIA Visual Recognition and Machine Learning Summer School in Paris. It attracted a total of 175 participants (31% from France, 50% from Europe and 20% from America and Asia). Next year the summer school will again be organized in Grenoble.