Etay Shefi

Spline-Based Wavelet Transforms on Non-Uniform Grids

Tuesday, July 14, 2015, at 15:30

Room 011, Kitot building, Faculty of Engineering.

~~
Speaker: Adi Dafni, 
M.Sc. student under the supervision of Prof. Shai Avidan and Prof. Yael Moses

Wednesday, July 15th, 2015 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering

Dynamic Object Segmentation in CrowdCam Images

Abstract

Segmenting moving objects is an essential tool in analyzing and visualizing dynamic scenes. Indeed many researches consider the task of extracting the dynamic objects mainly from video sequences. However, nowadays many dynamic events are captured by the observers by only a set of still images rather than videos. Moreover, no coordination between the photographers exists. We name such data crowd camera data or CrowdCam for short. Existing methods are not applicable for CrowdCam images since images are captured from different positions (can be wide-baseline) and at different times.
We address the problem of segmenting dynamic objects from a set of still images, taken by various un-calibrated cameras, with wide base line.
One of the main challenges is that dynamic regions (i.e, regions that are a projection of a dynamic object), are hard to detect, as they don’t obey any geometric constraint, and might undergo significant deformations. We take advantage of this attribute to distinguish them from the static regions, and propose a new approach that approximates the dynamic regions by elimination. Regions with low static probability in a given pair of images may either be dynamic or occluded. However, occluded regions are unlikely to be occluded with respect to all images. 
We compute a probability map in which the value of each pixel reflects its probability to belong to a static region, based on a combination of the confidence of correspondence along epipolar lines with respect to all images. Segments with low static probability are labeled dynamic.
The algorithm efficiently handles large displacements of the objects, and succeeds even when various sides of the objects are seen and when the objects are shaded in some of the images and highly illuminated in others. It works in a very broad setting, requires no prior knowledge about the scene, the camera characteristics or the camera locations.

~~Speaker: Ofir Levy, 
M.Sc. student under the supervision of Prof. Michael Margaliot and Prof. Lior Wolf

Monday, July 15th, 2015 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering

Live Repetition Counting

Abstract

The task of counting the number of repetitions of approximately the same action in an input video sequence is addressed. The proposed method runs online and not on the complete pre-captured video. It analyzes sequentially blocks of 20 non-consecutive frames. The cycle length within each block is evaluated using a convolutional neural network and the information is then integrated over time. The entropy of the network’s predictions is used in order to automatically start and stop the repetition counter and to select the appropriate time scale. Coupled with a region of interest detection mechanism, the method is robust enough to handle real world videos, even when the camera is moving. A unique property of our method is that it is shown to successfully train on entirely unrealistic data created by synthesizing moving random patches.