Working Student (all genders) – Robust Feed-Forward 3D Reconstruction for Dynamic Scenes

Feed-forward models for 3D reconstruction can directly predict scene geometry from images or videos without requiring computationally intensive scene-specific optimization. By combining large-scale pre-training, multi-view reasoning, and strong geometric priors, they represent an efficient alternative to classic reconstruction methods such as Structure-from-Motion, NeRF, and optimization-based 3D Gaussian Splatting.

Despite recent advancements, existing models still react sensitively to challenging real-world conditions. Occlusions, moving objects, changing lighting, night shots, reflections, rain, fog, and snow can lead to incomplete geometry, unreliable correspondences, and temporally inconsistent predictions. Higher robustness under such conditions is therefore of great importance, particularly for autonomous driving and robotic perception.

As a working student, you will support the development of robust feed-forward reconstruction models for dynamic scenes. You will investigate methods for handling occlusions, lighting changes, and adverse weather conditions. Furthermore, you will explore how large reconstruction models can be used as universal geometric backbones for downstream 3D scene understanding tasks, especially for Semantic Occupancy Prediction and 4D Occupancy Forecasting.

• Development and evaluation of feed-forward models for the reconstruction of dynamic 3D scenes from monocular or multi-view image sequences.
• Investigation of reconstruction robustness in the presence of partial and long-term occlusions, moving objects, and incomplete observations.
• Development of methods for improving geometric consistency under changing lighting, poor lighting conditions, shadows, and reflections.
• Evaluation and improvement of model performance under adverse weather conditions such as rain, fog, snow, and limited visibility.
• Comparison of the developed methods with relevant baselines, as well as documentation of technical and experimental results.

• You are currently pursuing a degree in Computer Science, Artificial Intelligence, Robotics, Electrical Engineering, Data Science, or a comparable field of study.
• You have very good programming skills in Python and practical experience with PyTorch.
• You possess a good understanding of Computer Vision, Deep Learning, 3D Geometry, or Multi-View Vision.
• Experience with depth estimation, optical flow, point clouds, camera pose estimation, NeRF, 3D Gaussian Splatting, or 3D reconstruction is a great advantage.
• Your language skills enable you to perform your role in English (minimum C1 level). German language skills are desirable, but not strictly required.

Within our uniform and transparent salary framework, the compensation for this position ranges between €15.50 and €19.50 per hour and is based on various factors such as qualifications and experience.

Daniela
+49 821 885882-0
Find Jobs in Germany on Arbeitnow