Florian Barthel

I am a PHD student at the Humboldt University of Berlin (Chair for Visual Computing) and a research assistant at Fraunhofer HHI (Vision and Imaging Technologies Group). My work focuses on synthesizing realistic 3D humans using state-of-the-art methods like 3D Gaussian Splatting, NeRF and 3D-aware Generative Adversarial Networks.

News

Publications

[CVPRW '24] Gaussian Splatting Decoder for 3D-aware Generative Adversarial Networks

In this work, we present a novel approach that combines the high rendering quality of NeRF-based 3D-aware Generative Adversarial Networks with the flexibility and computational advantages of 3DGS. By training a decoder that maps implicit NeRF representations to explicit 3D Gaussian Splatting attributes, we can integrate the representational diversity and quality of 3D GANs into the ecosystem of 3D Gaussian Splatting for the first time.

Paper
Project Page Code

[ECCV '24] Compact 3D Scene Representation via Self-Organizing Gaussian Grids

In this paper, we introduce a compact scene representation organizing the parameters of 3D Gaussian Splatting (3DGS) into a 2D grid with local homogeneity, ensuring a drastic reduction in storage requirements without compromising visual quality during rendering.

Paper
Project Page Code

[VISAPP '24] Multi-view Inversion for 3D-aware Generative Adversarial Networks

Our method builds on existing state-of-the-art 3D GAN inversion techniques to allow for consistent and simultaneous inversion of multiple views of the same subject. We employ a multi-latent extension to handle inconsistencies present in dynamic face videos to re-synthesize consistent 3D representations from the sequence.

Paper
Project Page Code

[ICMR '24] Creating Sorted Grid Layouts with Gradient-based Optimization

In this paper, we present a novel method for grid-based sorting that exploits gradient optimization for the first time. We introduce a novel loss function that balances two opposing goals: ensuring the generation of a "valid" permutation matrix, and optimizing the arrangement on the grid to reflect the similarity between vectors, inspired by metrics that assess the quality of sorted grids.

Paper
Slides

[ICCV '23] Haystack: a panoptic scene graph dataset to evaluate rare predicate classes

Haystack is a panoptic scene graph dataset that in contrast to existing scene graph datasets, includes explicit negative relation annotations. Negative relation annotations are important during evaluation, because they can drastically reduce label noise that occurs when relations are missed by annotators. During sampling, prior scene graph datasets will introduce some false negative labels, wheras Haystack guarantees that all negative relations are correct (assuming a perfect annotator).

Paper
Project Page Code

[SN Computer Science '22] Controlling 3D objects in 2D image synthesis

In this work, we propose a method that enforces explicit control over various attributes during the image generation process in a generative adversarial net. We propose a semi-supervised learning procedure that allows us to use a quantized approximation of object orientation for learning continuous object rotations. As a result, among many other attributes, our proposed method allows us to control object orientation in scenes that are rendered according to our specifications.

Paper

Projects

splatviz

This interactive viewer allows you to display and edit 3D Gaussian Splatting scenes in real time. By using a native python GUI library (imgui) we can directly manipulate the Gaussian python object just before rendering it. This enables endless editing and visualization possibilities.

GitHub

ERWIN (Student Project)

Erwin is a single-player puzzle game with co-op components. The idea is to play levels that would normally only be solvable in pairs, alone with the help of a time rewind mechanism.

itch.io

Thesis

2022 Master Thesis: Conditional Rendering of 3D Objects from 2D Images using Deep Generative Neural Networks

In this thesis, we will present a method that allows to conditionally synthesize photorealistic 3D-like 360° views of cars using multi-labeled 2D data.

PDF

2020 Bachelor Thesis: An Extensive Study of the StyleGAN Architecture for Deep Image Synthesis

This thesis analyzes the StyleGAN architecture, developed by the Nvidia research group Tero Karras et al. In the context of an ablation study, each component of the StyleGAN will be analyzed separately.

PDF