Markus Mathias



Incremental Object Discovery in Time-Varying Image Collections bibtex

T. Kontogianni, M. Mathias, B. Leibe

CVPR 2016

Object Discovery We address the problem of object discovery in time-varying, large-scale image collections. The content of large-scale image repositories is never static, as Millions of images are added to them each day, while others are withdrawn or deleted. Current object discovery algorithms do not yet address this issue. They typically operate in a static setting, making it necessary to re-run the entire clustering process whenever the underlying image database changes, even though only a small part of the clusters may be affected by the changes. We address the issue of growing databases by proposing a novel clustering method that allows for efficient reuse of the stored data and therefore allowing for efficient incremental updates of the database.


ATLAS: A Three-Layered Approach to Facade Parsing bibtex

M. Mathias, A. Martinovic, L. Van Gool

International Journal of Computer Vision (IJCV)

ATLAS We propose a novel approach for semantic segmentation of building façades. Similar to our previous work (ECCV 2012), the segmentation is performed in three levels of abstraction. In this work we revisit each of the abstraction layers by proposing various improvements and by providing extensive comparisons to other related work. We show improved results on the two datasets.


Face detection without bells and whistles bibtex

M. Mathias, R. Benenson, M. Pedersoli, L. Van Gool

Oral at ECCV 2014, 2.8% acceptance rate

Image 2 We present two surprising new face detection results: one using a vanilla deformable part model, and the other using only rigid templates (similar to Viola&Jones' detector). Both of these reach top performance, improving over commercial and research systems. We also discuss issues with existing evaluation benchmarks and propose an improved evaluation procedure.

Visit the for additional information.


Handling occlusions with franken-classifiers bibtex

M. Mathias, R. Benenson, R. Timofte, L. Van Gool

ICCV 2013

Image 2 Detecting partially occluded pedestrians is challenging.
Here we study why our classifier performs poorly under occlusion and we show that using a set of occlusion specific classifiers can significantly improve detection quality. Our proposed franken-classifiers obtain sub-linear grows in training and testing time, while keeping high detection quality.

Seeking the strongest rigid detector bibtex

R. Benenson*, M. Mathias*, T. Tuytelaars, L. Van Gool (* indicates equal contribution.)

CVPR 2013

Image 2 Previous work to boost detection quality explored using non-linear SVMs, more sophisticated features, geometric priors, motion information, deformable models, or deeper architectures. We use none of these.
>Using a single rigid classifier per candidate detection window we reach or improve state-of-the-art performance for pedestrian detection on INRIA, ETH and Caltech USA datasets.

Object detection for urban modeling (Object-detectie voor stadsmodellering) bibtex

M. Mathias

PhD Thesis 2013

Real world 3D city models are used for navigation, entertainment, virtual tourism and architecture. The creation of such models is a very demanding task that might take several man-years if performed manually. Automatic approaches using traditional computer vision techniques for 3D reconstruction work in a strictly bottom-up fashion by matching low-level information. The resulting quality of such approaches is wanting, especially for reflective, homogeneous or thin structures. On the other hand, top down information like the semantic structure of buildings can provide strong prior knowledge to be used in the reconstruction process. We incorporate such knowledge in the form of procedural building grammars and by semantically identifying several facade structures - such as windows and doors - in the input images by using object detectors. We show in two approaches that this leads to high quality building models, generated automatically.

In this doctoral thesis we will develop fast and robust object detectors that generalize well to different classes and present approaches to incorporate the information provided by detectors for facade modeling.

Traffic Sign Recognition - How far are we from the solution? bibtex

M. Mathias*, R. Timofte*, R. Benenson, L. Van Gool (* indicates equal contribution.)

IJCNN 2013

Image 2 We show that, without any application specific modification, existing methods for pedestrian detection and face recognition; can reach performances in the range of 95%∼99% of the perfect solution on current traffic sign datasets.


Fast stixels estimation for fast pedestrian detection bibtex

R. Benenson, M. Mathias, R. Timofte, L. Van Gool

ECCV 2012, CVVT workshop (Best paper award)

Image 2 We revisit the stixel computation method. Stixels computation now reaches 200 Hz (300 Hz latest version), and we can detect pedestrians at the record speed of 165 Hz (with state of the art quality).

A Three-Layered Approach to Facade Parsing bibtex

A. Martinovic, M. Mathias, J. Weissenberg, L. Van Gool

Oral at ECCV 2012, 3.6% acceptance rate

Image 2 We propose a novel approach for semantic segmentation of building façades. This segmentation is performed in three levels of abstraction, combining bottom-up labelling of superpixel with object detectors and top-down information coming from "weak architectural principles". We show state of the art semantic segmentation on two challenging datasets.

Pedestrian detection at 100 frames per second bibtex

R. Benenson, M. Mathias, R. Timofte, L. Van Gool

Oral at CVPR 2012, 2.5% acceptance rate

Image 2 We propose a new detector that improves both speed and quality over state-of-the-art single part detectors. We reach 50 Hz in monocular setup, and 135 Hz when using stixels on a street scene (including the stereo processing time).


Procedural 3D Building Reconstruction Using Shape Grammars and Detectors bibtex

M. Mathias, A. Martinovic, J. Weissenberg, L. Van Gool

3DIMPVT 2011

Image 2 We introduce a grammar-driven approach for reconstructing buildings and landmarks. We combine Structure-from-Motion and images-based analysis to perform inverse procedural modelling, shown for the example of Greek Doric temples.

Automatic Architectural Style Recognition bibtex

M. Mathias, A. Martinovic, J. Weissenberg, S. Haegler, L. Van Gool

3D-ARCH 2011

Image 2 When performing image based façade reconstruction, common pre-processing steps are usually performed manually, such as identifying if an image contains a façade or not, the alignment of horizontal and vertical structures to the image axis, façade splitting, and determination of the façade's style. We propose a full pipeline to address these pre-processing tasks.

Free CSS Templates by | Valid XHTML 1.0 | CSS 2.0