[cs-talks] UPDATED: PhD Proposal (Tue) + BUSec (Wed)
cs at bu.edu
Mon Mar 2 17:22:04 EST 2015
Probabilistic Frameworks for 3D Pose Estimation of Flying Animals in Variable-Resolution Multi-View Video Datasets
Mikhail Breslav, BU
(Updated) Tuesday, March 3, 2015 at 3:30pm in MCS 148
Abstract: Flying animals such as bats, birds, and moths, are actively studied by researchers wanting to better understand these animals' behavior and flight characteristics. Towards this goal, multi-view videos of flying animals have been recorded both in lab conditions and in natural habitats. The analysis of these videos has shifted over time from manual inspection by humans to more automated and quantitative approaches based on computer vision algorithms. In this thesis we study the largely unexplored problem of 3D pose estimation of flying animals in multi-view variable-resolution video data. This problem has received little attention in the computer vision community where flying animal datasets are uncommon. Additionally, existing solutions from researchers in the natural sciences have not taken advantage of advancements in computer vision research. Our thesis addresses this gap by introducing two probabilistic frameworks which build off of the rich computer vision literature on human pose estimation. These two frameworks follow a common approach of building a probabilistic model that leverages multiple camera views and temporal information, along with a prior capturing the 3D structure of the flying animal. The first framework uses a synthetic 3D graphics model along with a Markov Random Field to generate 3D pose estimates. The second framework, inspired by Pictorial Structures, learns to accurately detect body parts across multiple camera views. These body part detections are used to reconstruct 3D part positions leading to a 3D pose estimate. A post processing step finds the best sequence of 3D poses by considering temporal smoothness. In this thesis we also aim to compare and evaluate the two proposed frameworks on datasets obtained through collaboration with natural science researchers. Furthermore, we would like to make high quality multi-view flying animal datasets available with annotations to the greater computer vision research community.
Adaptively Secure Two-party Computation From Indistinguishability Obfuscation
Oxana Poburinnaya, BU
Wednesday, March 4, 2015 at 9:30 am in MCS 180 — Hariri Institute Seminar Room
Abstract: A basic challenge in the area of secure distributed computation is to achieve adaptive security, namely security against an adversary that can adaptively decide whom to corrupt during the execution of the protocol. Beyond providing better protection from realistic attacks than security against an adversary that controls a fixed-in-advance set of parties, adaptive security also provides strong resilience against leakage due to side channel attacks. However, all known general function evaluation protocols which provide full adaptive security have round complexity proportional to the circuit depth of the function. This is the case even with two-party protocols and even for honest-but-curious corruptions. We present the first two-round, two-party general function evaluation protocol that is secure against honest-but-curious adaptive corruption of both parties. In addition, the protocol is incoercible for one of the parties, and fully leakage tolerant. It requires a global (non-programmable) reference string and is based on one way functions and general-purpose indistinguishability obfuscation with sub-exponential security, as well as augmented non-committing encryption. A Byzantine version of the protocol, obtained by applying the CLOS compiler, achieves UC security with comparable efficiency parameters, but is no longer incoercible. The protocol uses Yao's garbled circuits and the Sahai-Waters puncturable deterministic encryption which allows embedding hidden triggers in a random-looking string. This is joint work with Ran Canetti and Shafi Goldwasser.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cs-talks