TUD Logo

TUD Startseite » ... » Künstliche Intelligenz » Neural Computation » Research

Neural Computation

Research areas


[1.  Segmentation with shape and shading]
[2.  Segmentation of multi-modal images]
[3.  Joint Non-rigid Motion Estimation and Segmentation]


1.  Segmentation with shape and shading

We propose a probabilistic segmentation scheme, which is widely applicable to some extend. Besides the segmentation itself our model incorporates object specific shape priors and shading. [more ]
orginal_lady_image segmented_lady_image


2.  Segmentation of multi-modal images

Assume two (or more) images of the same scene, but captured with different technologies (in the example below they are MRT- and PET- tomographies). A segmentation of these images is an interesting task itself even if all "other" parameters are known: parameters of the prior model, pixel correspondences etc. This is a nontrivial task if some segments are not distinguishable in one image at all. The key idea to overcome this difficulty is to segment the underlying scene rather then the images themselves. If, in addition, the images are not registered and furthermore some parameters of the prior model are unknown, then the task becomes really challenging. [more ]


3.  Joint Non-rigid Motion Estimation and Segmentation

Consider the following task: segment a non-rigidly moving object and estimate its motion field in a combined model. Let us assume that the object being compact, occludes the background but not itself as well as (for simplicity) still background. We model the task by a Markov Random Field where each pixel of the (reference) image is labelled by a combined segmentation/displacement label and assume a pixel-wise conditional independent p.d. of images given the hidden labelling. [more (pdf)]

hand_1 hand_2
hand1_ov hand2_ov
A hand moving in front of a plant: Two consecutive frames, segmentation in the first frame and deformed mesh in the second frame.

Stand: 1.2.2010, 14:28 Uhr
Autor: Dipl.-Inf. Denis Kirmizigül