Xinrui Lyu, MSc. EPFL in Electrical Engineering

"Problems worthy of attack prove their worth by fighting back." - Paul Erdos (1913-1996)

PhD Student

+41 44 632 23 74
ETH Zürich
Department of Computer Science
Biomedical Informatics Group Universitätsstrasse 6
CAB F52.1
8092 Zürich
CAB F52.1

I am a first-year Ph.D. student in Professor Gunnar Rätsch's group. My research interests lie on machine learning on clinical data and image processing.

Before joining the Biomedical Informatics Group at ETHZ, I received my M.Sc. in Electrical Engineering from École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, and my B.Eng. in Electronic Engineering from Tsinghua University, China. I have interned at Technicolor R&I Center, France where I worked on image searching algorithm for six months in 2015.

Abstract In this work, we propose a framework, dubbed Union-of-Subspaces SVM (US-SVM), to learn linear classifiers as sparse codes over a learned dictionary. In contrast to discriminative sparse coding with a learned dictionary, it is not the data but the classifiers that are sparsely encoded. Experiments in visual categorization demonstrate that, at training time, the joint learning of the classifiers and of the over-complete dictionary allows the discovery and sharing of mid-level attributes. The resulting classifiers further have a very compact representation in the learned dictionaries, offering substantial performance advantages over standard SVM classifiers for a fixed representation sparsity. This high degree of sparsity of our classifier also provides computational gains, especially in the presence of numerous classes. In addition, the learned atoms can help identify several intra-class modalities.

Authors Xinrui Lyu, Joaquin Zepeda and Patrick Perez

Submitted Proceedings of the British Machine Vision Conference (BMVC)

Link DOI

Abstract This paper presents an approach for using hierarchically structured multi-view features for mobile visual search. We utilize a graph model to describe the feature correspondences between multi-view images. To add features of images from new viewpoints, we designa level raising algorithm and the associated multi-view geometric verification, which are based on the properties of the hierarchical structure. With this approach, features from new viewpoints can be recursively added in an incremental fashion. Additionally, we designa query matching strategy which utilizes the advantage of the hierarchical structure. The experimental results show that our structure of the multi-view feature database can efficiently improve the performance of mobile visual search.

Authors X. Lyu, H. Li and M. Flierl

Submitted 2014 Data Compression Conference

Link DOI