Publications

2015
  • Amélie Royer, Christoph H. Lampert. "Classifier Adaptation at Prediction Time", IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

Internships

Summer 2014: Learning a prior for lifelong visual object categorization

Location : CV&ML Group at IST Austria, Vienna, Austria. Supervised by Christoph Lampert.

Abstract

When facing the task of classifying object images into categories, it can be useful to consider the context of the past seen queries to infer knowledge on the future inputs, rather than only using the immediate visual information of the object. In this work, we propose two combined algorithms bringing together a state-of-the-art image classifier and an online-learning system which gradually learns the intrinsic context of an input sequence.

To evaluate these algorithms, we design three methods to generate realistic sequences of queries; by “realistic”, we mean the object images are not uniformly sampled but rather part of a joint context and semantically related.
Our results establish that when dealing with such a realistic sequence of images, this combined visual-contextual approach outperforms the original classifier, by reducing the ambiguity on the classes.

Keywords: Image Classification, Domain Adaptation, Online Learning, Semantic Context Modelling, ImageNet.

Summer 2013: Event Retrieval in large video databases

Location : TEXMEX Team (INRIA Bretagne Atlantique, France).
Supervised by Hervé Jégou and Teddy Furon.

Abstract

This internship’s goal was to study the event retrieval task on large video databases: given a query video of a particular event, we want to retrieve every video of this same event in a database. The main issue is to find an accurate way to compare videos. Moreover, we need to represent our videos in a compact form, and keep reasonable computation times: we work on large video databases, thus a need of efficiency.

First, we study two approaches based on signal processing that were first introduced on the EVVE dataset ; the first one only uses the visual aspect of the videos, whereas the second one uses the chronological order of the frames in addition.
Next, we introduce a third event retrieval method, which uses a more compact description of our videos: it yields similar results to those of the previous approaches, with a benefit on memory usage and comparison time; however it requires fine tuning for the parameters. The three schemes were implemented using Matlab.

Keywords: Event retrieval, Video descriptors, Large video database

School Projects

Automatic verification of cryptographic protocols

Location : University of Rennes 1, 2013 - 2014. Supervised by Thomas Genet.
Joint work with: A. Chatalic, S. Kachanovich, L. Seguinot, B. Tessiau, A. Trieu.

Summary

Cryptographic protocols are nowadays used in many various aspects of our daily lives, which leads in many cases to a need of securing communications.

Some automatic tools are able to prove useful properties on these protocols; however they may be very difficult to understand and to use, due to a very particular formalism. Our goal is therefore to propose a user-friendly software overlay on one of this tools, in order to simplify the way it can be used.

This project is divided in two parts : our first step has been to study the state of the art in the domain of automatic cryptographic protocol verifiers; the second step is to implement the aforementioned overlay for the ProVerif verifier.

Keywords: Alice&Bob, Cryptographic protocols, ProVerif, Verification