Performed extensive research on the use of Visual Transformers for improving facial landmark detection, particularly occluded landmarks

Observed significant progress in identifying occluded landmarks due to the integration of self-attention in transformers, overcoming previous limitations of heatmap regression

Recorded a slight increase in error while processing clear facial images following the improvements in occluded landmark detection

A novel technique was developed to optimize over-parametrized neural networks.

Implementation of the approach identified an efficient subnetwork without needing extra training.

The method reduces computational expenses and memory demands without compromising performance and offers practical advantages across diverse neural network applications.

EyetRate utilized Magic Leap headset's eye-tracking cameras to estimate user's heart rate employing the Eulerian Video Magnification algorithm.

The project involved understanding the headset's functionalities, implementing the algorithm, and creating a software module to analyze the eye-tracking camera data.

Through extensive testing and refinement, color deviations induced by blood flow were accentuated to efficiently derive heart rate estimations, demonstrating the potential of eye-tracking technology in non-invasive health monitoring.

The research aimed to assess the feasibility of an MVP designed to enhance the efficiency of children's animation creation, with a primary objective of generating high-quality animations featuring a humanoid character from multiple input images.

Two different strategies were pursued: the first involved utilizing 2D image animation networks to reduce the time needed for animation creation.

The second method involved creating lifelike 3D avatars from multiple views and poses of the animated character, transforming static images into dynamic 3D animations with the help of motion-enabled software resources.