Previous | 
Home
Section: Partnerships and Cooperations
International Research Visitors
Internships
Aseem BEHL (from Nov 2012 to Dec 2012)
Subject: Optimizing Average Precision using Weakly Supervised Data. The average-precision support vector machine (AP-SVM) optimizes an upper bound on the average-precision (AP) loss, which is often used as a measure of accuracy for binary classification. However, it does not handle partially annotated datasets. To address this shortcoming of AP-SVM, we proposed a novel latent AP-SVM formulation, which allows us to learn an accurate set of classifier parameters by minimizing a carefully designed difference-of-convex upper bound on the AP loss.
Institution: International Institute of Information Technology (IIIT), Hyderabad (India).
Enzo FERRANTE (from June 2012 until October 2012)
Subject: Plane+Deformation 2D-3D multimodal data fusion. The goal of the internship was to study the use of discrete optimization methods in the context of 2D to 3D registration in biomedical image analysis. In particular the aim was to define a metric free graphical model formulation that is able to determine for a given 2D image the corresponding 3D volume plane along with the in plane deformation. The case of computer assisted surgery was considered as a test case between 2D interventional images and 3D pre-operative high resolution annotated data.
Institution: Universidad Nacional del Centro de la Provincia de Buenos Aires (Argentina)
-
Subject: Discriminative Parameter Estimation for Random Walks Segmentation. While random walks (RW) provide an efficient formulation for segmentation, there use is restricted by the lack of an accurate learning framework that estimates its parameters. The main difficulty is that a user can only provide a hard segmentation of a training sample, instead of the optimal probabilistic segmentation. We overcome this deficiency by treating the optimal probabilistic segmentation as latent variables, which allows us to employ the latent SVM formulation for parameter estimation.
Ishan MISRA (from May 2012 until Aug 2012)
Subject: Shape-from Shading analysis for Object Categories. The goal of the internship was to see whether shape-from-shading techniques can be used to recover the 3D geometry within an object category. Mr. Misra experimented with techniques for shape-from-shading under unknown illumination as well as surface recovery from a single image. Mr. Misra has delivered the source code for his software to our team, and we intend to use it in our on-going research.
Bharat SINGH (from May 2012 until September 2012)
Subject: Sub-space real-time Deformable Registration. The aim of this internship was to investigate the use of sub-space image representations towards defining an appropriate metric in the context of mono-modal and multi-modal fusion. Furthermore, it was studied their integration in a graph-theoretic framework for deformable fusion that can benefit from its implementation on modern parallel architectures like graphics processing units.
Eduard TRULLS (from January 2012 until April 2012)
Subject: Segmentation-Aware Image Descriptors. The goal of the internship was to construct appearance descriptors that can exploit segmentation information in order to achieve invariance to background changes. Mr. Trulls implemented a dense descriptor that uses soft segmentation masks, and demonstrated that this results in substantially more invariant descriptors; he evaluated these descriptors on image registration (optical flow) and wide-baseline matching (stereo) where state-of-the-art results were obtained. This work has been submitted for publication and is under evaluation.
Institution: Universidad Polytecnica de Catalunia (UPC) (Spain)
Visits to International Teams
Matthew BLASCHKO & Iasonas KOKKINOS (from June 2012 until August 2012)
Subject: Center for Language and Speech Processing: Towards a Detailed Understanding of Objects and Scenes in Natural Images Workshop. The objective of this workshop was to develop novel methods to reliably extract from images a diverse set of attributes, and to use them to improve the accuracy, informativeness, and interpretability of object models. The goal is to combine advances in discrete-continuous optimisation, machine learning, and computer vision, to significantly advance our understanding of visual attributes and produce new state-of-the-art methods for their extraction.