Technical Resume

Problem set: Fourier Transform

Implementation and exploration of the Discrete Fourier Transform. Demonstrating it's effectiveness for image compression.

Problem set: Projective Geometry, Corner Detection, Homographies

A problem set using Harris corner detection to demonstrate projective geometry. By detecting fixed locations on a wall and seeing the wall from different perspectives, we demonstrated transforming a simple image from one plane to another.

Problem set: Optical Flow & Motion Detection

A problem set using Hierarchical Lucas and Kanade for computing optical flow. Using this implementation, we estimated motion in a series of images and tried to interpolate intermediate frames.

Problem set: Particle Filters, Kalman Filters and Object Tracking

A problem set using particle and Kalman filters to track pedestrians in a series of videos. Needed to maintain track when camera was moving and after occlusions.

Problem set: Object Detection

A problem set implementing face recognition using principal component analysis (PCA), (Ada)Boosting and the Viola-Jones algorithm. Investigated the pros & cons of the various techniques and inherent difficulty of face detection.

Panorama

Align and stitch together a series of images into a panorama. Assumes we were given corresponding keypoints between 3 images. This was an assignment to learn about homographies, warping and image blending.

High Dynamic Range Imaging

This assignment was about learning the core algorithms behind computing HDR images based on the paper Recovering High Dynamic Range Radiance Maps from Photographs by Debevec & Malik.

Video Textures

This assignment was about creating infinitely looping videos by analyzing a video's texture. Based on the paper Video Textures by Schödl, A., Szeliski, R., Salesin, D., & Essa, I.

Content Aware Seam Carving

This goal of this project was to replicate the research results of the paper Seam Carving for Content-Aware Image Resizing by Shai Avidan, Ariel Shamir and Improved Seam Carving for Video Retargeting by Michael Rubinstein, Ariel Shamir. This involved understandiing and implementing forward seam carving, "backwards energy" and fine-tuning parameters.

Object Removal

The goal of this project was to replicate the research results of the paper Object Removal by Exemplar-Based Inpainting by Criminisi, A., Perez, P. & Toyama, K. Just like the Google Pixel 'magic erase' feature, our results demonstrated selecting some object in the image, masking it off and using the exemplar-based inpainting algorithm from the paper to fill in the pixels where the removed object used to be.

Supervised Learning

Exploration of supervised learning techniques. This assignment was not about implementing algorithms but tuning their parameters and analyzing what they can tell us about a dataset.

Learning algorithms analyzed: Decision tree (with pruning), Neural Networks, Boosting, Support Vector Machines (SVMs), k-Nearest Neighbors (KNN).

Used 2 classification problems to analyze the algorithms: Spambase dataset, which contains a database of email attributes and whether or not they are spam messages. And a Kaggle Spotify dataset which contains a collection of 1.2 million songs and their attributes like name, popularity, key, loudness, instrumentalness, tempo etc.

The goal was to create a binary classification problem for each dataset. For the spam dataset, whether or not some email was spam based on it's attributes. For the Spotify dataset, whether some song would be a hit for this artist (more popular than their average song).

Randomized Optimization

Exploration of random search techniques. Again this purpose of this assignment was not to implement the algorithms but instead to tune and analyze how they behave.

The optimization algorithms I chose were Genetic, Simulated Annealing, Randomized Hill Climb and MIMIC. They were applied to the k-colors problem, n-queens problem and the traveling salesman problem.

We needed to highlight advantages of different techniques on different optimization problems. For example, showing how one problem highlights advantages of genetic algorithms (traveling salesman), another advantages simulated annealing (n-queens) and another MIMIC (k-colors, sort of).

We also showed how Genetic, Simulated Annealing and Randomized Hill Climb can be used to find good weights for a neural network instead of backpropagation. We used the same dataset classification problem from the supervised learning project, SPAM filtering and song popularity.

Unsupervised Learning

Exploration of unsupervised learning algorithms: Clustering with k-means and expectation maximization. Dimensionality reduction through PCA, ICA and Randomized Projections. Finally, how dimensionality reduction can improve training a Neural Network.

Using the same SPAM and Spotify datasets as the other assignments.

Markov Decision Processes

Use OpenAi's gymnasium open source python library to explore Markov Decision Processes and how different algorithms approach different classes of MDPs. I chose to use the Frozen Lake (small, grid world problem) and the Taxi (larger grid world problem) to compare Value Iteration, Policy Iteration and Q-Learning as learning algorithms.