Ivan Laptev

email: <ivan.laptev at inria.fr>
phone: +33 (0)1 39 63 51 80
fax: +33 (0)1 39 63 79 87
address: INRIA, 23 avenue d'Italie
              75013 Paris, France  


News: INRIA CVML Summer School 2010 admission open

         Hollywood-2 dataset is available

         Open PhD position is available

         I moved to Paris, INRIA - Willow / ENS

I do research in computer vision. It is an exciting field combining several disciplines including mathematics, cognitive science, computer graphics and machine learning. I am particularly interested in video interpretation with the focus on visual motion analysis, event detection, object recognition and segmentation using adaptive low-level features, visual geometry and machine learning.

I am affiliated with INRIA Paris - Rocquencourt that is a unit of the French National Institute for Research in Computer Science and Control. I am with the WILLOW research group associated with École Normale Supérieure and leaded by Jean Ponce. I defended my PhD thesis at the Royal Institute of Technology (KTH) where I was at the Computer Vision and Active Perception laboratory (CVAP).

 
 
 

 

 

 

Projects:
 

The goal of this work is to recognize realistic human actions in unconstrained videos such as in feature films, sitcoms, or news segments. Our contributions concern (i) automatic collection of realistic samples of human actions from movies based on movie scripts; (ii) automatic learning and recognition of complex action classes using space-time interest points and a multi-channel SVM classifier (iii) Improved results for human action recognition (we achieve 91.8%) on the public KTH actions dataset.

 
 

Learning, recognition and localization of human actions in realistic videos such as movies, TV news and home recordings. We focus on atomic actions such as "drinking", "smoking", "hand shaking" and demonstrate action detection in challenging realistic scenarios with substantial variation of actions in terms of subject appearance, motion, surrounding scenes, viewing angles and spatio-temporal extents.

 
 

Learning and efficient detection of object categories such as "horses", "bicycles", "motorcycles" and "cars". The method is based on discriminative learning of histogram support regions and demonstrates competitive performance on PASCAL VOC'05 and VOC'06 benchmarks.

 
 

Detection and segmentation of periodic motion in complex scenes with non-stationary background and motion parallax. The method exploits periodicity as a cue and makes no strong assumptions about the background. 3D periodic motion is considered to overcome view variations of moving objects over time.

 
 

Recognition of classes of human actions such as "running", "walking" and "hand clapping". The method exploits local action representation in terms of space-time interest points and does not rely on motion segmentation. Results are provided for KTH action database and for other complex scenes.

 
 

Detection of local events in video with distinct properties in space-time. The positions and the spatio-temporal descriptors of events are computed invariantly to scale and velocity transformations in video. Space-time interest points enable matching of corresponding space-time points across video sequences and may be used for video alignment and motion recognition.

 
> l> > l> /html> > l> /html> ble> /html> > l> /html> > > l> /html>