Christopher Bulla

Coordinate selection for affine invariant feature description

In this paper, we present a method for affine invariant feature description. Based on the gradient distribution of an image region we calculate two basis vectors defining an affine invariant coordinate system, used to normalize the image region. The estimated basis vectors are non-orthogonal and allow for a precise representation of the gradient distribution. The proposed method can be combined with any feature detector and descriptor. Its performance is evaluated on globally affine transformed as well as on real world images and compared to state of the art methods for affine invariant feature description. The observed results outperform the results obtained by the SIFT feature detector and are comparable to the results obtained by ASIFT while having less computational complexity and being more flexibly applicable in case of local affine modifications.

Robust multi-view reconstruction from quasi-dense point cloud and poisson surface mesh

A novel method for generating a reliable initial surface mesh of the interested scene from a quasi-dense point cloud is presented in this article. Given multiple images taken from different points of view, a robust quasi-dense point cloud is acquired by accumulative triangulation of the potentially matched feature points. In our proposed method, the feature points are detected with Harris-corner and SIFT detectors. SIFT and DAISY descriptors are used to describe the neighboring environment of the detected features and provide efficient matching possibilities. The proposed method can be parallelized almost at each phase, which makes it suitable for image datasets of different sizes. The accuracy of the proposed method is evaluated quantitatively and qualitatively on different types and sizes of image datasets.

Detection of false feature correspondences in feature based object detection systems

In this paper we present a method for the detection of wrong feature correspondences in a local feature based object detection system. Common visual objects in different images share not only similar local features but also a similar spatial layout of their features. We will utilize this fact in order to distinguish between correct and wrong feature correspondences. The spatial feature layout will be modeled through a Delaunay triangulation. This triangulation is used to find clusters of feature correspondences that follow the same affine transformation. The decision whether a correspondence is correct or wrong can than be made based on this clustering. Our method is independent from the number of common objects in the images and produces reliable results even in difficult scenarios. It can also be used if the number of wrong correspondences is much higher than the number of correct correspondences. Experiments on real and synthetically generated images demonstrate the good performance of our approach.

Invariance against local affine deformation for feature based object detection systems

In this paper, we present a method to increase invariance against affine deformations in feature based object detection systems. We use the gradient distribution of an image region to calculate two non-orthogonal basis vectors defining an affine invariant coordinate system, which is used to normalize the image region. The proposed method is an intermediate processing step subsequent to the feature detection and can be combined with any feature detector and descriptor combination. Its performance is evaluated on locally affine transformed as well as on real world images and compared to state of the art methods for affine invariant feature description. The observed results outperform the results obtained by SIFT, ASIFT or the Harris-Affine based feature normalization method, without introducing significant additional demands on the memory requirement or the computational complexity.

Orthogonal transform of SIFT descriptors and the effect on size and performance of Fisher Vectors in visual search

Fisher Vectors have shown great capability for visual search. Their main drawback is their high dimensionality. We propose several methods to reduce the size of the Fisher Vectors by applying different preprocessing steps and dimension reduction techniques to SIFT descriptors. Also, we investigate the effects of PCA and DCT transforms employed on SIFT descriptors and the resulting improvement for an image retrieval application. We show that DCT transformed SIFT descriptors show superiority over other methods and that dimension reduction to 32 dimensions is possible as well as a reduction of number of clusters used for Fisher Vector generation without too much loss in the mean average precision.

Detection of Low-textured Objects

In this paper, we present a descriptor architecture, SIFText, that combines texture, shape and color information in one descriptor. The respective descriptor parts are weighted according to the underlying image content, thus we are able to detect and locate low-textured objects in images without performance losses for textured objects. We furthermore present a matching strategy beside the frequently used nearest neighbor matching that has been especially designed for the proposed descriptor. Experiments on synthetically generated images show the improvement of our descriptor in comparison to the standard SIFT descriptor. We show that we are able to detect more features in non-textured regions, which facilitates an accurate detection of non-textured objects. We further show that the performance of our descriptor is comparable to the performance of the SIFT descriptor for textured objects.

Object detection in quantized feature space

In this paper, we present a method for the detection of objects in a quantized feature space. Quantizing the feature space is a preprocessing step to compact the amount of data in large scale image retrieval and classification applications. A drawback, compared to the use of non-quantized features, is the loss in the ability to precisely detect and localize common objects across the images. Our method can handle this limitation and is based on the frequently used Bag of Visual Keypoints representation in combination with a sliding window approach. Thereby, it does not need any knowledge about the objects. Experiments on real and synthetic images show the good performance of our approach.