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

2 requirements, read out of the advert rather than guessed from the job title:

• Python
• PyTorch

Also mentioned, not required: computer vision, deep learning, 3D geometry, multi-view vision, depth estimation, optical flow, point clouds, camera pose estimation, NeRF, 3D Gaussian Splatting, 3D reconstruction, English (C1), German. Worth having, but their absence is not what gets a CV filtered out.

Abstract Feed-forward 3D reconstruction models can recover scene geometry directly from images or videos without costly scene-specific optimization. By combining large-scale pre-training, multi-view reasoning, and strong geometric priors, these models provide an efficient alternative to traditional reconstruction pipelines such as Structure-from-Motion, NeRF, and optimization-based 3D Gaussian Splatting. Despite recent progress, current models remain sensitive to challenging real-world conditions. Occlusions, moving objects, illumination changes, nighttime scenes, reflections, rain, fog, and snow can result in incomplete geometry, unreliable correspondences, and temporally inconsistent predictions. Improving robustness under such conditions is essential 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 occlusion, changing illumination, and adverse weather, and explore how large reconstruction models can serve as general-purpose geometric backbones for downstream 3D scene understanding, particularly semantic occupancy prediction and 4D occupancy forecasting.

These tasks interest you Develop and evaluate feed-forward 3D reconstruction models for dynamic scenes using monocular or multi-view image sequences.

Investigate reconstruction robustness under partial and long-term occlusions, moving objects, and incomplete observations.

Develop methods to improve geometric consistency under illumination changes, low-light conditions, shadows, and reflections.

Evaluate and improve model performance under adverse weather conditions such as rain, fog, snow, and reduced visibility.

Compare the developed methods with relevant baselines and document technical and experimental results.