*TALK I :*

Title : Graph and Tree Kernels ( Learning with Kernel Project )
Speaker : Arnab Ghosh and Viveka Kulharia
Timing : 4:00 PM - 4:30 PM , 16 January 16
Day : Saturday


* : *In the real world similarity between graphs is an important measure
for several real life problems like classification of Proteins and Enzymes
even in the Social Network analysis whereby 2 social graphs might have to
be clustered by similar properties.
In the quest to determine the similarity between pairs of graphs in 2005
Borgwardt et al looked at the similarity between the graphs in terms
of shortest path kernels which was an improvement over the random walk
kernels over product graphs which had its limitations on tractability but in
2010 SVN Vishwanathan et al came up with a novel method to compute
random walk kernels up to any length and that became much faster in terms
of computation than the Shortest Path Kernel .We looked at the 2010 paper
and implemented the methods for computing the random walk kernel that were
used in the paper and also implemented the shortest path kernel as shown by
the 2005 paper and compared the running times and the kernels obtained from


Title :  Vectorization of Real World Sketches
Speaker : Akshay Masare
Timing : 4:30 PM - 5:00 PM , 16 January 16
Day : Saturday


* : *There are still many artists who prefer to first sketch out their
ideas with pen and paper before recreating them digitally. During the
pre-design phase as well, pen and paper is the most used way of sketching.
But, the problem with using pen and paper in product design is that it
captures only one view of the object, and viewing from multiple views is
not possible. There have been attempts of using multiple sketch views for
generating a 3- dimensional view of the object, but the difficulty that all
these methods faced was detecting the borders of the object.

Usually edge detectors are used to find complex shapes like junctions,
straight lines and curves. But, sketch tokens provide a novel approach to
learning these features and detecting edge-based mid-level features. The
approach given was to learn supervised mid-level information from human
labelled edges. As, there is no good dataset available for hatched images,
we create one. The hatched lines in the images are usually drawn by the
artists in response to the contour of the object surface and the way light
hits the surface. As extracting these characteristics from a 2 dimensional
image is not possible, we decided to generate hatch lines on 3 dimensional