<div dir="ltr"><div class="gmail_default" style="font-family:arial,helvetica,sans-serif">When:    <span class="gmail-aBn" tabindex="0"><span class="gmail-aQJ">Monday, February 6th at 1:00 pm</span></span><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br>Where:   TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526<br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br>Who:      Payman Yadollahpour, TTIC<br><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif">Title: Exploring and Exploiting Diversity for Image Segmentation<br><br></div><div class="gmail_default"><span style="font-family:arial,helvetica,sans-serif">Abstract:</span><br><div style="text-align:left;font-family:arial,helvetica,sans-serif">Semantic image segmentation is an important computer vision task that is difficult </div><div style="text-align:left;font-family:arial,helvetica,sans-serif">because it consists of both recognition and segmentation. The task is often cast</div><div style="text-align:left;font-family:arial,helvetica,sans-serif">as a structured output problem on an exponentially large output-space, which is</div><div style="text-align:left;font-family:arial,helvetica,sans-serif">typically modeled by a discrete probabilistic model. The best segmentation</div><div style="text-align:left;font-family:arial,helvetica,sans-serif">is found by inferring the Maximum a-Posteriori (MAP) solution over the output</div><div style="text-align:left;font-family:arial,helvetica,sans-serif">distribution defined by the model. Due to limitations in optimization, the model cannot be <br>arbitrarily complex. This leads to a trade-off: devise a more accurate model that <br>incorporates rich high-order interactions between image elements at the cost of <br>inaccurate and possibly intractable optimization OR leverage a tractable model which <br>produces less accurate MAP solutions but may contain high quality solutions as other <br>modes of its output distribution. </div><div style="text-align:left"><span style="font-family:arial,helvetica,sans-serif"><br></span></div><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">This thesis investigates the latter and presents a two</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">stage approach to semantic segmentation akin to cascade models and proposal</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">generation works. In the first stage a tractable probabilistic</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">model outputs a set of high probability segmentations from the underlying</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">distribution that are not just minor perturbations of each other. Critically the</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">output of this stage is a diverse set of plausible solutions and not just a</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">single one. The first-stage reduces the exponential space of solutions to just a</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">handful of segmentations. In the second stage, a discriminatively trained re-ranking </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">model selects the best segmentation from this set. The re-ranking stage can </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">use much more complex features than what could be tractably used in the </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">probabilistic model, allowing a better exploration of the solution space than possible </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">by simply producing the most probable solution from the probabilistic model. The </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">formulation of the first-stage is agnostic to the underlying model and </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">optimization algorithm, which makes it applicable to a wide-range of models and </div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">inference methods. </div></span><div style="text-align:left"><br></div><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">Evaluation of the approach on a number of semantic image segmentation</div></span><span style="font-family:arial,helvetica,sans-serif"><div style="text-align:left">benchmark datasets highlight its superiority over inferring the MAP solution. </div></span><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif">Thesis Advisor: Gregory Shakhnarovich, <a href="mailto:greg@ttic.edu" target="_blank">greg@ttic.edu</a></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div><div class="gmail_default" style="font-family:arial,helvetica,sans-serif"><br></div><div><div class="gmail_signature"><div dir="ltr"><div><div dir="ltr"><div><div dir="ltr"><div><div dir="ltr"><font face="arial, helvetica, sans-serif">Mary C. Marre</font><div><font face="arial, helvetica, sans-serif">Administrative Assistant</font></div><div><i><font face="arial, helvetica, sans-serif" color="#3d85c6"><b>Toyota Technological Institute</b></font></i></div><div><i><font face="arial, helvetica, sans-serif" color="#3d85c6">6045 S. Kenwood Avenue</font></i></div><div><i><font face="arial, helvetica, sans-serif" color="#3d85c6">Room 504</font></i></div><div><i><font face="arial, helvetica, sans-serif" color="#3d85c6">Chicago, IL  60637</font></i></div><div><i><font face="arial, helvetica, sans-serif">p:(773) 834-1757</font></i></div><div><i><font face="arial, helvetica, sans-serif">f: (773) 357-6970</font></i></div><div><b><i><a href="mailto:mmarre@ttic.edu" target="_blank"><font face="arial, helvetica, sans-serif">mmarre@ttic.edu</font></a></i></b></div></div></div></div></div></div></div></div></div></div>
</div>