BOLT ⚡: Bilateral Filtering and Octree Lightweight Technique for Fast and Parameter Free Point Cloud Upscaling

In this paper, we introduce BOLT (Bilateral filtering and Octree Lightweight Technique), a novel fast and parameter free method for upsampling point clouds. We leverage the structural efficiency of the octree data structure and detail preserving properties of the bilateral filter to achieve a fast and parameter free upsampling method. Unlike the current state-of-the-art methods, BOLT does not require any parameters, deep learning, fine tuning or any type of training making it a suitable candidate for real-time applications. BOLT understands the underlying structure of the point cloud by dividing the point cloud into a hierarchical octree structure. Empty children are filled in the octree and outliers are smoothed using a bilateral filter.

Full Document

Compubox Reimagined: Using Computer Vision to Analyze Boxing

In combat sports, fight judging has a history of being controversial and inaccurate. Accurate and consistent scoring of punches is one of the most essential aspects of scoring a bout. To address this challenge, we present a computer vision-based architecture. It uses a convolutional neural network to detect hits. A recurrent neural network is utilized to classify punches. All these components work together to create a holistic and objective punch tracker. The CNN for hit detection takes a video frame and detects and localizes punches. Subsequently, the RNN takes a sequence of poses extracted from each video or live stream frame and classifies the punch or defensive action.

Full Document

A Reinforcement Learning Approach to Making Decisions On An NBA Court Using A Text Based Basketball Environment

In this project, a deep Q learning was applied to make decisions on an NBA court (from an offensive perspective). Using NBA stats, a text-based basketball game envi- ronment was created. Given a set of player’s stats, defender stats, team stats and work- ing under the assumption that defence was played man-to-man, numerous episodes were simulated until the agent converged on an optimal policy for making decisions at each state. To achieve this, deep Q-Learning was used [9]. It was shown that high performing players followed the model closely, whereas low performing players tended to stray away from the model’s recommendations.

Full Document

Notes on How Transformers Work

A high level set of notes based on Nvidia's Deep Learning Institutes Transformers course that explain how transformers work and are deployed.

Full Document

A Review on the Various Benefits of Electrical Vehicles

In this review, we go over the various benefits of driving an electric vehicle. Specifically, the benefits that relate to the environment, public health and cost savings and give an indepth review of each. To accomplish this, we look at vehicle data recorded by car rental company ZEROCAR, as well as other trusted sources such as census data and other data sets offered by Statistics Canada. We were able to show the significant benefits to the environment, public health and cost savings using these data sources. This shows us that using an electric car instead of a gas car does indeed make a difference in these aspects, and is worth making the consideration to switch.

Full Document