Projects Last Projects

This project introduces a novel method for monocular 3D reconstruction, utilizing a deep learning model that leverages Gaussian Splatting in a unique way. Instead of traditional rendering, we use Gaussian Splatting as an efficient 3D latent representation for features within an autoencoder architecture. This approach prioritizes speed and minimal input data, enabling real-time applications on local devices.
Our pipeline defines a 3D latent representation that allows for memory and time-efficient rendering. By learning Gaussian parameters during training, we streamline the reconstruction process, bypassing the need for extensive input data and real-time computation. Extensive evaluations and ablations demonstrate that our model achieves robust and accurate 3D reconstructions from a single image input. This innovation addresses key challenges in the field, significantly enhancing reconstruction efficiency and speed without compromising accuracy.

Please, see project report.

Please, see final presentation.