Presentation Gallery
Applications of computer vision have seen great success recently, yet there are few approaches dealing with visual illustrations. We propose a collection of computer vision applications for parsing genetic models. Genetic models are a visual illustration often used in the biological sciences literature. These are used to demonstrate how a discovery fits into what is already known about a biological system. A system that determines the interactions present in a genetic model can be valuable to researchers studying such interactions. The proposed system contains three parts. First, a triplet network is deployed to decide whether or not a figure is a genetic model. Second, a popular object detection network YOLOvS is trained to locate regions of interest within genetic models using various deep learning training techniques. Lastly, we propose an algorithm that can infer the relationships between the pairs of genes or textual features present in the genetic model.
Library LinkWe propose a computer vision system for parsing gene model maps. Gene model maps are visual illustrations (often used in the biological sciences literature) that are used to demonstrate how a discovery fits into what is already known about a biological system. Furthermore, determining what interactions occur between a set of genes is valuable for researchers. Our work comprises three parts. First, we train a triplet network to decide whether or not a figure is a gene model map. Next, we use Google's Cloud Vision optical character recognition to extract the text of genes occurring within the gene model map. Secondly, we repurpose the architecture of YOLOv5 trained on our synthetic dataset of biological diagrams to find locations of genes and relationships. Lastly, we proposed a deep network that is able to infer the relationships between the gene pairs present in the gene model. Components of this system power the search functionality of Bio-Analytic Resource for Plant Biology tool at the University of Toronto.
PDFVideo has rapidly become one of the most common and largest sources of visual information. Although merely effortless to record large amounts of video data, raw videos often require significant editing until being best suited for viewing. User edited videos often only include segments of the raw video which are deemed to be interesting by the editor. State-of-the-art video summarization methods have accomplished generating high-quality summaries from raw videos. Attempting to beat current state-of-the-art video summarization methods is a daunting task. For this reason, my honours thesis focuses towards attempting to discover different new methods for video segmentation. In this honours thesis, I present my experiments using long short term memory (LSTM) cells to successfully detect boundaries in a video, and rate the interestingness score of the final frame, given a sequence of frames from a video. Using the TVSum50 dataset to conduct these experiments, I conclude that an LSTM based model can learn temporal dependencies within a sequence of frames to successfully detect boundaries and rate the Interestingness of a frame.
PDFVideo has rapidly become one of the most common and largest sources of visual information. This explosion of accessible video data through many mediums such as smartphones, video cameras, webcams, etc. has brought us to an age of big video data. Although merely effortless to record large amounts of video data, raw videos often require significant editing until being best suited for viewing. User edited videos often only include segments of the raw video which are deemed to be interesting by the editor. State-of-the-art video summarization methods have accomplished generating high-quality summaries from raw videos. How can we use frame-level features to improve the summarization by changing the velocity of the video at different segments?
PDFThe UOIT Registrar’s Office reached out to the Vialab to build a to visualize student retention data for exploratory data analysis with an emphasis on patterns which are predictive of student withdrawal. The RetentionVis tool relies on user interaction through application of filters to the visualizations to assist the Registrar's Office to answer the question: ‘Why are students dropping out?’. All post-secondary intuitions' main source of revenue is tuition from undergraduate studies. Student retention is key to the success of any institution, not only for revenue but for their reputation. The dashboard is an online interactive tool created in JavaScript’s D3 library hosted by UOIT. The tool requires the user to make various selections including: faculties/programs, years, timeslots, and GPA ranges of interest with the goal of finding opportunities for improvement at UOIT.
PDF