Learning 3d Representations in Function Space from 2d Images

In this project, we have learned 3D representations in a function space from 2D renders. We have implemented and experimented with the methods proposed in the academic paper: [Occupancy Networks - Learning 3D Reconstruction in Function Space] as well as some ideas from the blog post: [Implicit Decoder Part 1: 3D Reconstruction].

We have used a preprocessed dataset from [Implicit Decoder], which contains the following data:

• pixels: 2D renders from the 3D models. There are 24 renders for each 3D model.

• points: Randomly sampled points from the 3D space.

• values: The ground truth values for the occupancy of the voxel that represents the corresponding point.

• voxels: The 3D models in voxel representation.

Using the occupancy net, we were able to create a 3d render of an object for just a 2d image.

Please navigate to the colab notebook [Here] to have a look at the code and a more in-depth explanation.

Emotion Detection and Expression Editing using Variational Autoencoders and Latent Space Arithmetic

The goal of this project is to classify the emotion expressed by a person based on their facial features in an image. While expressions are usually good indicators of emotions, they are not always accurate. Any bias that may be observed in the model is unintended.

We have also experimented with blending images of different emotions to create new ones, which are exciting and interesting. These are achieved using variational autoencoders

To have a look at the code please follow [this] link to the colab notebook

Spotify Music Recommendations using Sparse Overcomplete Autoencoders

A recommender system, or a recommendation system, is a type of filtering system that seeks to predict the “rating” or “preference” a user would give to an item.

Recommender systems are typically classified into the following categories:

• Content-based filtering

• Collaborative filtering

• Hybrid systems

This project uses content-based filtering to make music recommendations to the user after receiving five songs as per their preference.

The aim of this project is to:

1. Acquire the metadata, audio features data, and lyrics of Billboard Hot 100 (henceforth referred as BBHOT100) tracks for each year in the time period 1960 to 2021. This is accomplished by web scraping and Spotify API.

2. Perform extensive exploratory data analysis on the processed BBHOT100 dataset to derive insights and see how music preferences have evolved over time.

3. Generate a content-based music recommendation system based on sparse overcomplete autoencoders and deploy the system on Streamlit using Streamlit Cloud.

4. Develop a XGBoost model to predict rank of a track and further study it using SHAP values to increase transparency and interpretability of the model.

Please navigate to the GitHub repo [here] for have a look at a more comprehensive summary of the project