Some projects I've done during my time at UCL (still ongoing, updates to come). The code is present, as well as either pdf reports, slides or documented Jupyter notebooks to explain what has been done:
-
My Master's thesis will soon be uploaded. It is still being reviewed. Topic: "Super-resolution and modality synthesis for tumour-impaired brain MRI scans from the Africa brain archive".
-
COMP0169 Machine Learning for Visual Computing: coding a neural network from scratch, using linear and non-linear classifiers, building with Pytorch autoencoders and VAE using in particular convolutional layers
-
COMP0026 Image Processing: implementing segmentation algorithms (K-means, tresholding), a coding a method for warping two images (rigid transform with a triangulation, and meshless), and implementing Poisson image editing (from "Poisson Image Editing", Pérez et al., SIGGRAPH 2003) for image interpolation (seamless cloning, mixing gradients, and texture flattening).
-
COMP0137 Machine Vision: implementing an expectation-maximization model to separate a mixture of Gaussian for visual data, and a coursework on tracking (with a particle filter) and homographies with the aim of producing a 3D cube in augmented reality.
-
COMP0114 Inverse Problems in Imaging: investigating optimization solutions depending on the norm chosen, the SVD decomposition of well and ill-conditioned matrices, comparing algorithms using a matrix VS using a handle function, implementing a Kyrlov solver for denoising blurry images using the Laplacian of the image, and conducting a paper review of "Stable signal recovery from incomplete and inaccurate measurements" from Candès et al., 2005. Final coursework on the Radon transform and wavelet denoising, with further experiments on inpainting in sinogram space (methods ranging from the isotropic Laplacian inpainting to the anisotropic Total Variation inpainting).
-
COMP0118 Computational Modelling for Biomedical Imaging: the first coursework covered parameter mapping, uncertainty estimation, and model selection on part of the Human Connectome Project data. The second coursework dealt with statistics, especially on F-tests, p-values, estimation and error spaces. The last coursework is a mini research project on working with MRI data for the placenta (rather than the brain studied previously). The report was written like a research paper. Extensive testing of models, comparison of the models using BIC, making computational/quantitative tests to obtain a qualitative conclusion.
-
COMP0119 Acquisition and Processing of 3D Geometry: the first coursework covered ICP, point-to-point VS point-to-plane, the importance of the mesh's initial position, its robustness to noise, and global alignment. The second coursework dealt with mesh denoising and smoothing, using methods involving the Laplace-Beltrami operator (Laplacian for 3D meshes), uniform and non-uniform Laplacian discretizations, modal analysis, implicit and explicit smoothing, and finally a differential geometry part dealing with the first and second fundamental forms of a 3D figure. Final project on recreating and assessing the DeepICP network from Lu and al (2019): using pre-trained models, creating a synthetic dataset, analysing and criticising a research paper.
-
(Not appearing) COMP0130 Robot Vision and Navigation: creating scripts for GNSS and Dead Reckoning position estimation, modifying existing code from the drivebot and g2o libraries, and ORB-SLAM2 to experiment and assess the performance of each algorithm's parameter.
-
(Not appearing) COMP0027 Computer Graphics: creating simple scenes in GLSL. BRDF textures, anti-aliasing, ray-tracing, rasterisation, the rendering equations