3D Neural Scene Representations for View-Invariant State Representation Learning

My course project for MIT's Advances in Computer Vision class.

Installation

Clone the repo and execute the following commands from the repository's root.

Create a virtual environment:

python -m venv state_encoder_3d_env

Activate the environment:

source state_encoder_3d_env/bin/activate

Install the state_encoder_3d package in development mode:

pip install -e .

Training and Evaluation

Training data can be generated using scripts/generate_planar_cube_data.py. All training, evaluation, and visualization scripts can be found in the scripts/ folder.

Results

A detailed report can be found here.

The below figure is taken from the report and compares the performance of different models.