Section: New Results

Matching/Segmentation

Participants : Haithem Boussaid, Iasonas Kokkinos, Chaohui Wang, Bo Xiang, Ahmet Besbes, Ben Glocker, Nikos Komodakis, Nikos Paragios.

• Rapid Deformable Part Model Detection: in [27] we introduce a Branch-and-Bound technique which efficiently finds the most promising configuration of a pictorial structure model given an image. The fastest previously known techniques are linear in the image size; our technique has a best-case complexity that is logarithmic in the image size. When evaluated on standard datasets (Pascal benchmark) our technique gives a 5- to 15-fold speedup. Moreover, when evaluated in the multi-object detection problem our technique's complexity scales sublinearly also in the number of objects, resulting in 20- to 100- fold speedups when evaluated with 20 object categories.

• Segmentation with Deformable Graph-based Priors: in [22] we have introduced a novel formulation to address deformable segmentation using graph-based priors while being able to handle partial-correspondences. Segmentation was formulated as a matching task, where candidate correspondences were determined using boosting, and the assignment problem was solved using MAP inference constrained by a graph-based deformable prior. The notion of missing/erroneous correspondences was introduced in the process leading to state-of-the art results once compared with prior art in the field. The same prior was used in the context of the segmentation of tagging MR heart images [37] . The main contribution of this paper was the exact estimation of the region-based probability likelihood within a pair-wise MRF through the use of Stokes theorem and integral images.

• Deformable Model-based 3D reconstruction: in [23] we introduce a model-based optimization approach to the 3D reconstruction of Femur images using a small set of low-dose X-Ray images. We use a parametric deformable model of the Femur surface and fit it to the acquired data by optimizing its parameters. We incorporate in our optimization criterion multiple aspects of the problem, namely the 3D surface- to 2D plane projection, region-based statistics, and edge-based terms. Our evaluation includes both in vitro and in vivo experiments, where our method is shown to yield promising results, while alleviating the need for time demanding, manual annotations.

• Pose-invariant Higher Order Graph-based Priors: in [36] we have introduced a novel method for 3D model inference from 2D images in the absence of camera pose parameters. The method exploits higher (fourth) order priors, which alleviate the need of the estimation of the camera parameters. Furthermore, the proposed formulation couples 3D model inference with 2D correspondences and results on a single shot solution for both problems in the absence of knowledge of the observer internal and external parameters.