Research Areas

Research Areas

Probabilistic models of low-level vision

High-order MRFs

We study probabilistic models of photographic images, as well as those of other dense scene attributes, with the aim of capturing their key spatial statistics accurately. A particular focus of our work is on high-order random field models, such as Fields of Experts, that capture the spatial properties of images in extended pixel neighborhoods and can be trained on example data. For these models we also study probabilistic and deterministic inference algorithms. Finally, we work on their applications to problems such as image restoration, including image denoising and image deblurring.

Image motion estimation

Scene Flow
3D Scene Flow

Our research on image motion aims to provide robust techniques for accurate 2D and 3D motion estimation from image sequences (optical flow and 3D scene flow). On one hand, we study advanced spatial models of image motion, including approaches based on statistical analyses and learning. On the other hand, we research various aspects of motion estimation algorithms based on such models. Finally, we also work on benchmarking motion estimation algorithms.

Human tracking and pose estimation

Pose Estimation
3D Pose Estimation

We are interested in detecting and tracking humans in still images and video sequences. Moreover, we aim to estimate the persons' body pose, either in 2D or in 3D, from monocular images and videos. Our emphasis is on techniques that cope with the variability of human appearance in real world conditions, as well as the complexity of realistic scenarios, where often multiple people are present simultaneously.

Object recognition and scene understanding

Traffic scene understanding
Traffic scene understanding

Beyond humans, we are also working on detecting objects from a broader range of object categories. One aspect of our work is to address appearance variations of objects as they occur in many realistic scenarios. This includes weakly-supervised part-based object models, as well as learning appearance representations from data. Another aspect of our work is to incorporate context from the 3D scene or other objects present toward more integrated approaches of scene understanding.