Conducted thorough research on the implementation of Visual Transformers to enhance facial landmark detection, with a specific focus on optimizing the identification of occluded landmarks. Previously impeded by the heatmap regression approach, the transformers exhibited a notable improvement in effectively discerning occluded landmarks through the integration of self-attention. However, a consequent increase in error was observed when processing unobstructed facial images.

A novel stochastic gate-based pruning technique was devised for the purpose of efficiently optimizing over-parametrized neural networks. Through the implementation of this approach, a subnetwork capable of attaining equivalent performance to the target network was successfully identified, without the need for supplementary training. This research presents a highly promising opportunity to substantially diminish computational expenses and memory demands, whilst simultaneously upholding exemplary performance standards. Moreover, these advancements hold considerable practical advantages across diverse domains that rely on the utilization of neural networks.

EyeRate was a self-initiated endeavor, leveraging the eye-tracking cameras of the Magic Leap headset, to compute the user's heart rate employing the Eulerian Video Magnification algorithm. The undertaking encompassed comprehending the headset's functionalities, executing the algorithm, and constructing a software module to analyze the eye-tracking camera data. Through rigorous testing and meticulous refinement, the project adeptly derived heart rate estimations by accentuating color deviations induced by blood flow. As evidenced by its accomplishments, EyeRate effectively demonstrated the immense potential of eye-tracking technology in facilitating non-invasive health monitoring.

I conducted this research to assess the feasibility of an MVP aimed at enhancing the efficiency of children's animation creation. The primary objective was to generate high-quality animations featuring a humanoid character by utilizing multiple input images. Two distinct approaches were explored: the first focused on leveraging 2D image animation networks to reduce animation creation time, while the second delved into the realm of 3D avatar creation. This exploration involved generating lifelike 3D avatars from multi-view and multi-pose images of the animated character, utilizing techniques that seamlessly transform static images into dynamic 3D animations with the support of motion-enabling software resources available in the market.