Research Projects

Multi-object tracking

Our work focuses on improving the data association step in multiple object tracking by exploiting contextual information (social or spatial), improving the detection matching score with a siamese network architecture or merging multiple inputs (head detections or superpixels) in a single optimization problem.

Image-based localization

Deep Learning for 6D pose estimation? Not so fast! We show an experimental comparison of SIFT-based vs CNN-based methods, propose a new CNN+LSTM architecture to improve localization accuracy and a new challenging TUM LSI dataset with ground truth from a laser scanner.

Video Object Segmentation

One-Shot Video Object Segmentation (OSVOS), a fully-convolutional neural network architecture that is able to successively transfer generic semantic information, learned on ImageNet, to the task of foreground segmentation. With just a glimpse of the first frame, we segment an object in a full video! SOA on DAVIS!

MOTChallenge. The Multiple Object Tracking Benchmark.

A new standardized way to compare multiple people tracking algorithms. Provided detections, ground truth for the training sequences, a set of challenging test sequences. Three challenges have been launched so far since 2014! A new pedestrian detection challenge is also here.

Past projects

Learning Proximal Operators

We propose to replace the proximal operator of the regularization used in many convex energy minimization algorithms by a denoising neural network. The latter therefore serves as an implicit natural image prior, while the data term can still be chosen independently. We obtain state-of-the-art reconstruction results, indicating the high generalizability of our approach and a reduction of the need for problem-specific training.