Eduard Trulls
Computer Vision lab @ EPFL
Address: EPFL-IC-CVLab, BC 300, Station 14, CH-1015 Lausanne (Switzerland)
E-mail: eduard dot trulls at epfl dot ch / Tel: +41 216 93 81 95

I am currently a post-doc at the Computer Vision Lab at EPFL in Lausanne, Switzerland, working on Deep Learning topics under the supervision of Prof. Pascal Fua. I obtained my PhD from the Institute of Robotics in Barcelona, Spain. My thesis explored novel strategies to enhance local, low-level features (e.g. SIFT, HOG) with global, mid-level data such as motion and segmentation cues, and was co-advised by Francesc Moreno and Alberto Sanfeliu. Before my PhD I worked in mobile robotics.

Links: GitHub / LinkedIn / Google Scholar / Videos



Note: The top three computer vision conferences (CVPR/ICCV/ECCV) are highly competitive, with low acceptance rates: 20-30%.
You may also be interested in the Google Scholar Metrics.

LF-Net: Learning Local Features from Images
Y. Ono, E. Trulls, P. Fua, K.M. Yi
arXiv:1805.09662, 2018

We introduce an novel architecture and training scheme to learn a deep network for local feature extraction from scratch, from whole images. To do so we exploit camera pose and depth cues which we show can be generated with standard, automated SfM frameworks, without further supervision. We show how to train our network effectively from two images, in a Siamese-like architecture, by breaking differentiability in one branch to create a virtual target. Our approach outperforms the state of the art while running in real time for QVGA images.

Learning to Find Good Correspondences
K.M. Yi(*), E. Trulls(*), Y. Ono, M. Salzmann, V. Lepetit, P. Fua (*: equal contribution)
Conference on Computer Vision and Pattern Recognition (CVPR), 2018 (oral presentation: 2.1% rate)
code / poster / bibref

We develop a deep architecture to learn to find good correspondences for wide-baseline stereo. Our solution is based on putative keypoint matches, which we learn to label as inliers or outliers while simultaneously using them to recover the relative pose, as encoded by the essential matrix. Our solution is simple (no convolutional or fully-connected layers), small (4 Mb), easy to train (state of the art matching outdoors scenes with only 59 images) and generalizes well, particularly in contrast to image-based dense methods.

LIFT: Learned Invariant Feature Transform
K.M. Yi(*), E. Trulls(*), V. Lepetit, P. Fua (*: equal contribution)
European Conference on Computer Vision (ECCV), 2016
code (theano) / code (tf) / poster / video / supplemental / bibref

We introduce a novel, Deep Network architecture that implements the full feature extraction pipeline: keypoint detection, orientation estimation, and feature description, with a single architecture that can be trained end-to-end while preserving differentiability. Our models outperform state-of-the-art methods on multiple benchmarks.

Learning to Match Aerial Images with Deep Attentive Architectures
H. Altwaijry, E. Trulls, J. Hays, P. Fua, S. Belongie
Conference on Computer Vision and Pattern Recognition (CVPR), 2016
poster / bibref

We present a novel approach to match ultra-wide-baseline aerial images with a deep convolutional neural network. In addition to the binary classification task (match/no match) we show that local correspondences can still be of use, by incorporating an attention mechanism that produces a set of probable matching regions. Our models outperform the state of the art and approach human accuracy on a challenging problem.

Dense Segmentation-aware Descriptors
E. Trulls, I. Kokkinos, A. Sanfeliu and F. Moreno-Noguer
Chapter in Dense Image Correspondences for Computer Vision, Eds. C. Liu and T. Hassner, Springer, 2015

We wrote a chapter for a book on dense matching, extending the work presented in our CVPR 2013 paper. We pursue invariance to occlusions and background changes by introducing segmentation information within dense feature construction. The core idea is to use the segmentation cues to downplay the features coming from image areas that are unlikely to belong to the same region as the feature point.

Discriminative Learning of Deep Convolutional Feature Point Descriptors
E. Simo-Serra(*), E. Trulls(*), L. Ferraz, I. Kokkinos, P. Fua and F. Moreno-Noguer (*: equal contribution)
International Conference on Computer Vision (ICCV), 2015
code / poster / supplemental / spotlight / bibref

We present a novel framework to learn discriminative local image descriptors with a siamese architecture of deep Convolutional Neural Networks, trained with pairs of corresponding and non-corresponding patches. We deal with the large number of potential pairs with the combination of a stochastic sampling of the training set and an aggressive mining strategy biased towards patches that are hard to classify. By using the L2 distance during both training and testing we obtain descriptors which can be used as a drop-in replacement for SIFT. We demonstrate consistent performance gains over the state of the art and very good generalization properties.

Enhancing low-level features with mid-level cues
E. Trulls
PhD thesis, 2015

We use global cues to enhance local features with mid-level information. In particular, we present novel techniques to enrich features such as SIFT or HOG with motion or segmentation cues, and demonstrate consistent improvements on multiple problems, including dense matching and object detection.

Segmentation-aware Deformable Part Models
E. Trulls, S. Tsogkas, I. Kokkinos, A. Sanfeliu, F. Moreno-Noguer
Conference on Computer Vision and Pattern Recognition (CVPR), 2014
poster / spotlight / bibref

We combine bottom-up segmentation (SLIC superpixels) with DPMs. We use the superpixels to build soft segmentation masks at every scale and position. We use the masks to "clean up" the HOG features, splitting them into foreground and background channels.

Dense segmentation-aware descriptors
E. Trulls, I. Kokkinos, A. Sanfeliu, F. Moreno-Noguer
Conference on Computer Vision and Pattern Recognition (CVPR), 2013
code / poster / spotlight / bibref / site

We exploit segmentation data to construct appearance descriptors that can deal with occlusions and background motion. We use the segmentation to build soft segmentation masks, and downplay measurements likely to belong to a different region. We integrate this with SIFT and also with SID, a dense descriptor invariant by design to rotation and scaling.

Spatiotemporal descriptor for wide-baseline stereo reconstruction of non-rigid and ambiguous scenes
E. Trulls, A. Sanfeliu, F. Moreno-Noguer
European Conference on Computer Vision (ECCV), 2012
poster / spotlight / bibref / site

We use temporal consistency to match appearance descriptors and apply it to stereo on very ambiguous video sequences. Previous works define descriptors over spatiotemporal volumes, which is not applicable to wide-baseline scenarios—instead we extend 2D descriptors with optical flow estimates to capture the change around a feature point in time.

Autonomous navigation for mobile service robots in urban pedestrian environments
E. Trulls, A. Corominas Murtra, J. Pérez-Ibarz, G. Ferrer, D. Vasquez, Josep M. Mirats-Tur, A. Sanfeliu
Journal of Field Robotics, 2011
bibref / site

An extension of our 2010 IROS paper. We switch from 2D to 3D data for localization, and present experiments in a new urban area: a street open to the general public in the city of Barcelona, Spain.

Efficient use of 3D environment models for mobile robot simulation and localization
A. Corominas Murtra, E. Trulls, J. M. Mirats Tur, A. Sanfeliu
International Conference on Simulation, Modelling, and Programming for Autonomous Robots (SIMPAR), 2010. Also in Simulation, Modelling, and Programming for Autonomous Robots, Lecture Notes in Computer Science, 2010.

This paper provides a detailed description of a set of algorithms to efficiently manipulate 3D models to compute physical constraints and range observation models, used for real-time robot localization.

Autonomous navigation for urban service mobile robots
A. Corominas Murtra, E. Trulls, O. Sandoval, J. Perez, D. Vasquez, J. M. Mirats Tur, M. Ferrer, A. Sanfeliu
International Conference on Intelligent Robots and Systems (IROS), 2010

We present a solution for fully autonomous navigation on urban, pedestrian environments, designed for highly mobile robots based on Segway platforms.

3D Mapping for Urban Service Robots
R. Valencia-Carreño, E. Teniente, E. Trulls, J. Andrade-Cetto
International Conference on Intelligent Robots and Systems (IROS), 2009

We present an approach to build 3D maps from 3D range data as the main input, based on the probabilistic alignment of the point clouds using SLAM.

Combination of Distributed Camera Network and Laser-based 3D Mapping for Urban Service Robotics
J. Andrade-Cetto, A. Ortega, E. Teniente, E. Trulls, R. Valencia, A. Sanfeliu
Workshop on Network Robot Systems, International Conference on Intelligent Robots and Systems (IROS), 2009

An overview of the URUS project.



For information, supplemental material and videos:


Some of the projects I have been part of: