PhD Defense by Fausto Milletari

Speaker: Fausto Milletari Date: Thursday, February 8, 2018 Time: 15:00 Location: FMI-Building, Room 00.12.019

Hough Voting Strategies for Segmentation, Detection and Tracking

Abstract:

Object detection, segmentation and visual tracking are extremely important problems in both computer vision and medical image analysis. Most of the recent scientific efforts focused on proposing machine learning based approaches that can tackle and solve these problems appropriately. Methods that are based on handcrafted features, the so called shallow approaches, have been widely used and explored until very recently and were often employing machine learning algorithms such as boosting, support vector machines, random forests coupled with a careful choice of manually engineered features which were designed in a specific manner for each task. Most recently these approaches have been replaced by deep learning methods which are as well capable of learning features directly from raw data and can capture semantically meaningful information in a hierarchical and structured fashion. Such approaches, which are articularly suited for vision tasks, deliver in some cases superhuman performances when applied to challenging problems. Although machine learning approaches as such delivered outstanding performances on a number of challenging tasks, many methods -- especially in the field of medical image analysis -- cannot still be applied in a straightforward manner. The lack of large amounts of annotated training data, the presence of noise and artifacts, the low inter-class versus the high intra-class variability of the samples, and other domain-specific factors, often limit the performances of the models. In my talk I will show how voting strategies can be used to tackle detection, segmentation and pose estimation problems relying on voting strategies which look only at image parts and assemble the resulting knowledge into a global decision. This approach overcomes the limitation of current machine learning methods in all those cases where, due to the nature of the data and despite appropriate training, the uncertainty of the decision over previously unseen data remains high.