Adhark Incorporation (Sept 2017 - Present)
Role: Data Scientist
Project: Image analysis with deep learning, to evaluate the visual preferences of different audience groups
Deep learning platform: Tensorflow
Programming Language: Python
Unsupervised Deep Feature Extraction for Visual Complexity Analysis
In this work, we propose a new direction in unsupervised information extraction from
intermediate convolutional layers of deep neural networks. We focus on visual com-
plexity, an image attribute that humans can subjectively evaluate based on the level
of detail in the image. We derive an activation energy (AE) metric that combines
convolutional layer activations to quantify visual complexity. In addition, we introduce
Savoias, a visual complexity dataset that compromises of more than 1,400 images
from seven diverse image categories.
ExerciseCheck: Remote Monitoring and Evaluation Platform for Home Based Physical Therapy
We present ExerciseCheck, a remote monitoring and evaluation platform for individuals involved in a home exercise program. The goal of the platform is to give patients feedback about their performance and, if needed, and how they should adjust their movements. ExerciseCheck is designed for a therapist to remotely monitor a patient in real time, enabling the therapist to give instant feedback or further instructions.
Dynamic Adjustment of Physical Exercises Using Reinforcement Learning
In this project, we propose a system for divculty adjustment of exercises in physical
therapy. We use pattern recognition algorithms (Dynamic Time Warping and Spectral
Arc Length), to extract performance measures; then we design our online reinforcement
learning algorithm to dynamically adjusts the divculty conguration of exercises based
on the users performance.
Posture Modeling for Exercise Assessment
We propose a method to capture and model the postures of a user for during a physical
therapy exercise. We capture the skeletal data using the Microsoft Kinect, extract
features and train a Gaussian Mixture Model (GMM) to represent postures of the user
during an exercise. Our model enables autonomous detection of incorrect postures
during an exercise.