In this project, we have optimized a network for both liver and liver metastases automatic segmentation purposes. We have investigated several different network architectures for the segmentation task. First, we tested vanilla UNet to have an idea as our baseline model. Next, in order to optimize the performance, we supplanted the UNet encoder backbone with networks from the VGG family. We used UNet with the VGG19 backbone, and then we experimented with the VGG19 backbone pretrained on ImageNet. Further extensive experimentation on heterogeneous inter dataset transductive transfer learning was done to generate segmentation for datasets without ground-truth. This work was done in laboratoire de traitement de l’information médicale (LaTIM), Brest, France. [ For report and code please send me an email ]

A year-long Tutored Research and Development Project (TRDP), focusing on Interpreting and comparing Convolutional Neural Networks (CNNs). As CNNs are black-box systems, we are investigating what these kinds of networks are actually learning at a neuron level. In other words, what visual latent features are being extracted by each convolutional units or a group of units at a certain layer, and what effects do they have on the ultimate discriminative power of the network. We are doing so by interpreting the neurons in terms of known human recognizable visual concepts. This work was done in The Video Processing and Understanding Lab (VPU) and University of Bordeaux. [ Report and Code]

The objective of the work is twofold. First, Estimating the optical flow and global motion with RANSAC algorithm. Second, Measure the information content derived from the video sequences with different metrics such as Mean squared error(MSE), Entropy and Peak signal-to-noise ratio (PSNR) for study. Later a comparative analysis was carried out between the response of these metrics for a set of video sequences containing different kinds of motion in the sequences. Also, the effects of hyper-parameter deltaT have been explored. [ Report and Code ]

This is a python console application that generates mesh (faces) from a pgm3d file using only the NumPy package. It takes input pgm3d file then processes the voxel values into bins aka labels. Afterward, generate the mesh or faces between voxels with different labels and save it as a .obj file. You can view the .obj file by importing them into Mesh lab (https://www.meshlab.net/#download) [ Code and output samples ]

The objective of this work is to detect moving objects on a stationary and dynamic background. In the beginning, a basic method was created with the blind and selective update for the grayscale images, which was later remodelled to support colour images. Right after that, a routine was designed to filter out stationary objects that were appearing and disappearing in the frame. In the next stage of development, a shadow removal technique was introduced to mitigate the effect of shadows in the foreground mask. Further, advanced background subtraction algorithms such as Gaussian Mixture Models were explored in search of improvement and as well as to accommodate segregation of background and foreground in dynamic background situations. [ Report and for Code ]

The goal of this notebook is to implement a tomographic reconstruction sinograms with filtered back projection algorithm. The sinogram were acquired with different projection geometry. First, The reconstrction for parallel beam projections and cone beam projections were implmented for 2D singoram. Then reconstruction for 3D cone beam projection has also been implemented. [ Code ]

Object detection and identification is a very challenging task in video-surveillance systems. Among multiple approaches and methods, this report is solely focused on the method based on foreground segmentation which can be divided into two sub-tasks i.e., blob extraction and blob classification. In the blob extraction stage, OpenCV’s floodfill module was used while in the classification stage a statistical classifier was implemented. Another challenge faced in video sequence analysis is to detect stationary foreground objects. For this reason, a routine was developed to detect stationary foreground blobs in the frame based on foreground history images. Further, a snippet of MATLAB code was generated to form an improved classification model. [ Report and for Code ]

we explored colour-based and gradient-based histogram tracking at first. Generally, both models have their strengths and weaknesses depending on the application scenario. We combined those methods to get the best of both worlds. Further, We developed a tracking model based on the Kalman filter using different velocity and acceleration models. Finally, we tested our models in the real-world scenario and attempted to reach an optimal set of parameters for the best performance for each video sequence. [ Report 01, Report 02 and for Code 01 , Code 02 ]

The main object of the project was to develop an IoT-based home automation system that can be controlled and monitored by a UWP (Universal Windows Platform) application and Voice. Also, the UWP application can be deployed to any windows device and the application has an external API to be integrated with the cloud. [ Report and for code please send me an email ]