Ground Truth People:
Katrin Honauer, Karsten Krispin, Alexander Brock, Oliver ZendelAbstract:
Consider you would like to solve an image processing problem. You know that researchers have developed hundreds of algorithms which are supposed to work for you. Which one do you choose? Unfortunately, only very few evaluation papers have been published for many research areas and even fewer implementations exist. One of the main problems for algorithm evaluations is the large variance of possible input data sets in computer vision and image processing. Our research is dedicated to create meaningful and comprehensive datasets annotated with ground truth for image processing applications.
- Realistic Synthetic Ground Truth
In this project we evaluate the effects of physically plausible renderings of ground truth sequences for optical flow estimation. Therefore, we create controlled real sequences and reconstruct these using raytracing techniques. Comparisons between optical flow estimates of real scenes and synthetic scenes show that the results are similar both in mean endpoint error and its spatial distribution. Based on these insights, we carry out studies focusing on the usefulness of synthetic ground truth sequences.
- Real Data With Approximate Ground Truth
Together with Robert Bosch AG we built a portable camera system which is capable of continuously recording image sequences with very high spatio-temporal resolution. These sequences are then analyzed with state-of-the-art algorithms developed at the HCI in order to generate approximate ground truth data. This data is then used to verify the outcome of other algorithms for performance evaluations.
Light Field People:
Hendrik Schilling, Marcel Gutsche, Hamza Aziz Ahmad und Maximilian Diebold (PostDoc)Abstract:
Feature methods play an important role in a wide range of computer vision, image understanding and visual inspection applications. HCI covers this research area from two different perspectives: The mathematical and algorithmic investigation of robust feature algorithms on the theoretic side and the research and implementation of novel physically based imaging techniques to capture object properties (features).
Both approaches are largely intersecting, affect each other strongly and are targeted to provide novel robust feature method for real world applications.
- Local Descriptors for 3D Vector Fields
We introduced a set of local descriptors for the analysis of 3D vector field data. Our methods are based on a novel harmonic (frequency) representation of local spherical vector field patches, which allows an efficient and robust computation of rotationally invariant descriptors as well as a fast normalized cross-correlations (over all possible 3D rotations) and accurate rotation offset estimation between such patches.
- 3D Object Retrieval
In this project, we investigated an adaptation the Bag of Features (BoF) concept to 3D shape retrieval problems. The BoF approach has recently become one of the most popular methods in 2D image retrieval. We extent this approach from 2D images to 3D shapes. Following the BoF outline, we addressed the necessary modifications for the 3D extension and presented novel solutions for the parameterization of 3D patches, a 3D rotation invariant similarity measure for these patches and a method for the codebook generation.
- Learning of Optimal Illumination for Material Classification
Surface defects can be regarded as deviations in reflectance and/or geometry. Given an illumination series, the characteristics of surface defects are learned and propositions about optimal lighting environments are made.