Section: New Results

Shape, Grouping and Recognition

In [30] we have extended our prior work on dense scale- and rotation- invariant image descriptors to take into account soft segmentation information. This allows us to discard measurements stemming from background structures, and as such renders our descriptors invariant to background changes and occlusions. This has allowed us to obtain state-of-the-art results on tasks such as large-displacement optical flow and wide-baseline stereo. We have made the implementation of these descriptors publicly available.

In [22] we have started exploring the potential of combinatorial optimization in the medical imaging realm. We cast the problem of finding a 3D structure (a brain tumor) as that of finding the mode of a nonparametric distribution, constructed through Kernel Density Estimation. Current techniques for doing this (e.g. Mean Shift mode-seeking, Fast Gauss Transforms, etc.) are either iterative, or linear in the number of pixels, with a typically large constant. Instead, we develop a scheme that involves a very low-constant linear-time preprocessing step, and then uses Branch-and-Bound for fast mode estimation. As such it is scalable to large volumes, and serves as a rapid initialization of a region segmentation algorithm.

In [17] we pursue a Reinforcement Learning-based approach to couple image observations with a grammar-based method to partitioning a building facade. For this we expressed 2D grammar- based image parsing as a Markov decision process where an agent has to take actions in an environment so as to maximize some notion of cumulative reward (reflecting the segmentation quality). This allowed us to accelerate previous stochastic hill-climbing approaches to image parsing by more than an order of magnitude.

In [27] we extended our previous work on fast object detection by developing a sparse-coding method for the efficient sharing of computation among multiple object models. In particular the first processing step of `part score' computation was originally performed separate per object category; instead, we propose to do it `in batch mode', so as to exploit the commonalities that exist among object parts. Building on recent developments in sparse coding we have managed to construct a compact basis for this task, which in the end gave us a two-fold acceleration over our previous fastest algorithms.