This repository explores Autoencoding Gaussian Splats for 2D image representation. Gaussian splatting has emerged as a powerful technique for modeling images and 3D scenes using parameterized Gaussians. In this project, we apply autoencoders to learn compact representations of images modeled as 2D Gaussian splats.

1. Construct a dataset of trained Gaussian splats using CIFAR-10.
2. Design and train an autoencoder to encode and reconstruct Gaussian splat representations.
3. Compare Gaussian splat autoencoding with traditional pixel-based autoencoding.

• We use gsplat to fit Gaussians to CIFAR-10 images, storing parameters like position, scale, rotation, opacity, and color.
• Different autoencoder architectures (deep, convolutional, ResNet-based) are explored for encoding Gaussian splats.
• Experiments analyze compression efficiency, reconstruction quality, and feature disentanglement.

.
├── configs/              # Configuration files for experiments
├── constants/            # Constants and transformation utilities
├── data/                 # Dataset scripts and preprocessed data
├── images/               # Visualization results (e.g., loss curves, comparisons)
├── logs/                 # Logs from different model training runs
├── models/               # Implementations of autoencoders and trainers
├── references/           # Reference implementations and utilities
├── report/               # LaTeX report files and compiled PDF
├── results/              # Experimental results and analysis
├── slurm/                # SLURM batch scripts for job scheduling
├── style/                # Custom matplotlib styles
├── submodules/           # External repositories (e.g., ResNet-18 autoencoder)
├── tests/                # Jupyter notebooks for experimenting with different setups
├── utils/                # Utility functions for data processing and visualization
├── example.ipynb         # Provided example from mentors
├── LICENSE               # License file
├── README.md             # Project documentation
└── requirements.txt      # Python dependencies

1. Clone the repository:

   git clone https://github.com/mokot/visual-representation-learning.git
   cd visual-representation-learning

2. Install dependencies:

   pip install -r requirements.txt

3. Run experiments: Refer to tests/ folder for Jupyter notebooks with different experimental setups.

• gsplat: Open-source library for Gaussian splatting (github.com)
• CIFAR-10 Dataset: Standard benchmark dataset for image representation.

This project is licensed under the MIT License.

Both authors contributed equally to the conceptualization, research, and implementation of this project.

• Rok Mokotar (LMU Munich) – Rok.Mokotar@campus.lmu.de
• Federico Bernardo Harjes Ruiloba (LMU Munich) – f.harjes@campus.lmu.de