NOTE: I have moved to the Vision Group at Queen Mary University of London. My new page is located HERE.
Shape detection with nearest neighbour contour fragments
We present a novel method for shape detection in natural scenes based on incomplete contour fragments and nearest neighbour search. In contrast to popular methods which employ sliding windows, chamfer matching and SVMs, we characterise each contour fragment by a local descriptor and perform a fast nearest-neighbour search to find similar fragments in the training set. Based on this idea, we show how to learn robust object models from training images, to generate reliable object hypotheses, and to verify them.
BIMP: Fast, repeatable keypoints based on the visual cortex
Biologically inspired processing is often slow due to repeated convolutions. We have managed to create a biologically-inspired keypoint detection algorithm which is both fast and competitive with the state of the art in terms of repeatability. Efficient CPU and GPU-based implementations are available.
Fast, learning free object detection based on Naive Bayes Nearest Neighbours
We designed a fast Bayesian algorithm for category-level object detection in natural images. We modified the popular Naive Bayes Nearest Neighbour classification algorithm to make it suitable for evaluating multiple sub-regions in an image, and offered 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. An implementation based on OpenCV is available.
Context-based probabilistic scene interpretation
General scene understanding remains one of the great unsolved problems in Computer Vision and Artificial Intelligence. Building on deep Bayesian Compositional Hierarchies (BCH), we present a general framework for probabilistic scene understanding based on conceptual models and aggregate hierarchies. We show how uncertain evidence can be integrated into a stepwise interpretation process and how top-down context can resolve uncertainty in detection. Results are presented in the domain of facade images.
Knowledge-based behaviour recognition
The objective of this project was to provide the methods and techniques that enable construction of vision systems that can perform task oriented categorization and recognition of objects and events in the context of an embodied agent. Our SCENIC system used knowledge-based representation built on taxonomical and partonomical relations and constraints between objects and events. The system is able to recognise complex actions, disambiguate uncertain visual evidence, and make predictions about future events.
Segmentation of 3D volumes
The aim of this project was to develop methods for analysing 3D tomography images of medium density fibreboards (MDF), a material made of wood commonly used in the industry. The tasks included segmentation of individual fibres, finding the contact surface between the fibres, determining the amount of adhesive resin, calculating the lumen volume and visualising the results.