Research Projects

• Modelling hippocampal place cells for visual localization

  Biologically inspired method that uses neural networks to improve the visual recall of already visited places

  Keywords: Biologically inspired algorithms, localization, neural networks, regression, machine learning, pattern recognition, big data, open databases.

  More information soon…

• Visualizing visual path data with t-SNE

  Bag-of-visual-words high dimensional data visualization using t-distributed stochastic neighbor embedding (t-SNE)

  Visual path data from RSM dataset (http://rsm.bicv.org) in the form of appearance-based BOWs entails a high dimensional spaces. Concretely, the dimension of this space will be the choice of dictionary size for the bag-of-words model. Prior to the advent of deep learning, dictionaries of size 256, 400 or 4,000 were typically found in the literature.

  In this project, I have made use of t-distributed stochastic neighbour embedding (t-SNE) (van der Maaten and Hinton, 2008) to visualize the different points as members of different spaces or corridors within the RSM dataset.

  More info soon…

• Associating locations from wearable cameras

  Evaluation pipeline of context based image retrieval methods when used to provide visual localization from wearable cameras

  In this work we suggest an evaluation methodology that estimates the error distributions in inferred position with respect to a ground truth. We assess and compare standard approaches from the field of image retrieval, such as SIFT and HOG3D, to establish associations between frames.

  Keywords: Localization, open datasets, evaluation scripts, pipeline, context-based image retrieval, machine learning, multi-class classification, SVM, kernel-trick, distance metrics, bag-of-words, encoding methods.

  More information soon…

• RSM dataset

  We define a visual path as the video sequence captured by a moving person in executing a journey along a particular physical path. More information…

  Keywords: Visual localization, open databases, big data.

• Hand-held object recognition benchmark

  Categorization benchmark of popular object recognition algorithms when tested against the SHORT dataset

  In this project I evaluated the performance of a number of popular object recognition algorithms, with differing levels of complexity, when tested against SHORT. One of the main motivations was to devise the open challenges in the visual object recognition of objects that are being held by users and to establish a baseline for SHORT to be compared against other datasets.

  Benchmarked methods: SIFT descriptor matching, SIFT bag of words (BOW) model with different encoding techniques: hard assignment (HA), locality constrained linear coding (LLC) and Fisher Vectors.

  Benchmarked datasets: VOC 2007, Caltech-101, SHORT.

  Keywords: Wearable cameras, assistive context, open databases, object recognition and retrieval, classification, SVM, Fisher Vectors, VLAD, GMM, VLFEAT, descriptors, bag-of-words, encoding methods, open source code.

  More information soon…

• SHORT-100 dataset

  SHORT dataset for hand-held object recognition from mobile or wearable cameras

  The SHORT-100 dataset of grocery products is an updated and practical dataset for studying object recognition and retrieval in the challenging scenarios of hand-held objects and mobile or wearable cameras, and with emphasis in the assistive case for blind and partially sighted users. More information …

  Keywords: Assistive context, open databases, object recognition and retrieval, classification, groceries.

• Pairwise descriptor comparisons for localization

  Pilot study of image matching techniques when used in the context of visual localization.

  With the early versions of the RSM dataset, I conducted studies of the performance of descriptor matching techniques when used with indoor paths sequences to answer the questions a) in which corridor am I on? and b) where am I along the corridor?

• Features for the visual biopsy of polyps

  Development of image processing techniques to extract visual features for the automatic classification of polyps in the bowel.

  Awarded `A’ mark. The project was accomplished in collaboration with medical image processing company Medicsight and NHS expert colonoscopists.This project applied image processing, computer vision and machine learning techniques to endoscopic images, with the purpose of improving the characterisation and classification of large bowel polyps in real-time during colonoscopy.

  I improved the classification accuracy on two datasets by a 2% and 5% respectively (up to 86.44% and 94.64%) by adding four new features to the existing five.

  I also designed and implemented different approaches to narrow the margin of classification error by producing measurements of feature stability such as feature variability across the surface of polyps.

  I used several classification techniques, most importantly support vector machines (SVMs) and linear discriminant analysis (LDA).

  Download thesis