Kasim Terzić

Microstructural analysis of lignocellulosic fiber networks

The structure of wood based medium density fiberboard (MDF) has been studied using synchrotron radiation-based x-ray microtomography (SRμCT.) Fully automated 3D segmentation and analysis routines have been developed in order to gain information about individual fibers, the distribution of the fiber material, fiber orientation, fiber surfaces and size and location of contact areas. Representative samples of the analyzed volume data are presented to demonstrate the results of the implemented methods using the VIGRA image processing library.

High-Level Expectations for Low-Level Image Processing

Scene interpretation systems are often conceived as extensions of low-level image analysis with bottom-up processing for high-level interpretations. In this contribution we show how a generic high-level interpretation system can generate hypotheses and initiate feedback in terms of top-down controlled low-level image analysis. Experimental results are reported about the recognition of structures in building facades.

Fast cortical keypoints for real-time object recognition

Best-performing object recognition algorithms employ a large number features extracted on a dense grid, so they are too slow for real-time and active vision. In this paper we present a fast cortical keypoint detector for extracting meaningful points from images. It is competitive with state-of-the-art detectors and particularly well-suited for tasks such as object recognition. We show that by using these points we can achieve state-of-the-art categorization results in a fraction of the time required by competing algorithms.

BIMP: A real-time biological model of multi-scale keypoint detection in V1

Abstract We present an improved, biologically inspired and multiscale keypoint operator. Models of single- and double-stopped hypercomplex cells in area V1 of the mammalian visual cortex are used to detect stable points of high complexity at multiple scales. Keypoints represent line and edge crossings, junctions and terminations at fine scales, and blobs at coarse scales. They are detected by applying first and second derivatives to responses of complex cells in combination with two inhibition schemes to suppress responses along lines and edges. A number of optimisations make our new algorithm much faster than previous biologically inspired models, achieving real-time performance on modern GPUs and competitive speeds on CPUs. In this paper we show that the keypoints exhibit state-of-the-art repeatability in standardised benchmarks, often yielding best-in-class performance. This makes them interesting both in biological models and as a useful detector in practice. We also show that keypoints can be used as a data selection step, significantly reducing the complexity in state-of-the-art object categorisation.

Biological models for active vision: towards a unified architecture

Building a general-purpose, real-time active vision system completely based on biological models is a great challenge. We apply a number of biologically plausible algorithms which address different aspects of vision, such as edge and keypoint detection, feature extraction, optical flow and disparity, shape detection, object recognition and scene modelling into a complete system. We present some of the experiments from our ongoing work, where our system leverages a combination of algorithms to solve complex tasks.

Context-aware classification for incremental scene interpretation

Appearance-based classification is a difficult task in many domains due to ambiguous evidence. Knowledge about the relationships between objects in the scene can help resolve this problem. In this paper, we present a new probabilistic classification framework based on the cooperation of decision trees and Bayesian Compositional Hierarchies, and show that introducing contextual knowledge in the form of dynamic priors significantly improves classification performance in the façade domain.

Methods for reducing visual discomfort in stereoscopic 3D

Visual discomfort is a significant obstacle to the wider use of stereoscopic 3D displays. Many studies have identified the most common causes of discomfort, and a rich body of literature has emerged in recent years with proposed technological and algorithmic solutions. In this paper, we present the first comprehensive review of available image processing methods for reducing discomfort in stereoscopic images and videos. This review covers improved acquisition, disparity re-mapping, adaptive blur, crosstalk cancellation and motion adaptation, as well as improvements in display technology. HighlightsVisual discomfort in stereoscopic 3D viewing is an obstacle to adoption of 3D media.This paper presents the first review of literature on reducing discomfort.The review spans numerous fields and references over 200 publications.We include a discussion of future directions and impact of new technologies.

Context-Based Probabilistic Scene Interpretation

In high-level scene interpretation, it is useful to exploit the evolving probabilistic context for stepwise interpretation decisions. We present a new approach based on a general probabilistic framework and beam search for exploring alternative interpretations. As probabilistic scene models, we propose Bayesian Compositional Hierarchies (BCHs) which provide object-centered representations of compositional hierarchies and efficient evidence-based updates. It is shown that a BCH can be used to represent the evolving context during stepwise scene interpretation and can be combined with low-level image analysis to provide dynamic priors for object classification, improving classification and interpretation. Experimental results are presented illustrating the feasibility of the approach for the interpretation of facade images.

An efficient Naive Bayes approach to category-level object detection

We present a fast Bayesian algorithm for category-level object detection in natural images. We modify the popular Naive Bayes Nearest Neighbour classification algorithm to make it suitable for evaluating multiple sub-regions in an image, and offer a fast, filtering-based alternative to the multi-scale sliding window approach. Our algorithm is example-based and requires no learning. Tests on standard datasets and robotic scenarios show competitive detection rates and real-time performance of our algorithm.

Real-Time Object Recognition Based on Cortical Multi-scale Keypoints

In recent years, a large number of impressive object categorisation algorithms have surfaced, both computational and biologically motivated. While results on standardised benchmarks are impressive, very few of the best-performing algorithms took run-time performance into account, rendering most of them useless for real-time active vision scenarios such as cognitive robots. In this paper, we combine cortical keypoints based on primate area V1 with a state-of-the-art nearest neighbour classifier, and show that such a system can approach state-of-the-art categorisation performance while meeting the real-time constraint.