Englische Zusammenfassung
The project adapts pre-trained 3D foundation models to rotating-drum granular-flow video, using video preprocessing, model specialization strategies and a joint spatiotemporal learning head. Students evaluate 3D reconstruction and material-property inference against experimental baselines and conduct ablation studies. The work includes documenting the adaptation pipeline and releasing reproducible inference and training code; 100% remote work is possible.
Überblick
- Projektdetails
- Video-Driven World Modeling for Granular Powder Flow Dynamics
Textstelle · Seite 1
- “# Video-Driven World Modeling for Granular Powder Flow Dynamics” Seite 1
- Organisation
- Im PDF nicht angegeben
- Projektformate
- IDP
Textstelle · Seite 1
- “Interdisciplinary project (IDP)” Seite 1
- IDP
- Studienabschlüsse
- Im PDF nicht angegeben
- Projektziel
- Specialize pre-trained 3D foundation models for granular powder flow, jointly recovering temporally consistent 3D powder-bed geometry and inferring physical material properties from monocular video.
Textstelle · Seite 1
- “Given monocular video of evolving powder surfaces, our model will simultaneously solve two coupled tasks: (1) recover temporally consistent 3D geometry of the powder bed; (2) infer implicit physical characteristics governing the observed flow dynamics.” Seite 1
Themen und Aufgaben
- Themen
- granular powder flow dynamics
Textstelle · Seite 1
- “Video sequences of granular powder flow carry coupled signals of geometry, motion and intrinsic material physics.” Seite 1
- 3D computer vision and reconstruction
Textstelle · Seite 1
- “The core idea is to leverage the geometric and visual priors encapsulated in off-the-shelf 3D foundation models” Seite 1
- granular material physics
Textstelle · Seite 1
- “material properties such as cohesion, friction and flow rheology” Seite 1
- 3D foundation model adaptation
Textstelle · Seite 1
- “This project focuses on specializing pre-trained 3D foundation models for this physical granular system.” Seite 1
- granular powder flow dynamics
- Anwendungsbereiche
- rotating-drum granular flow experiments
Textstelle · Seite 1
- “Curate and preprocess monocular video sequences from rotating-drum granular flow experiments:” Seite 1
- rotating-drum granular flow experiments
- Methoden und Werkzeuge
- monocular video sequences
Textstelle · Seite 1
- “Given monocular video of evolving powder surfaces” Seite 1
- pre-trained 3D foundation models
Textstelle · Seite 1
- “specializing pre-trained 3D foundation models for this physical granular system” Seite 1
- frame sampling
Textstelle · Seite 1
- “frame sampling, ROI cropping, photometric normalization, camera calibration and pseudo-multi-view construction” Seite 1
- ROI cropping
Textstelle · Seite 1
- “ROI cropping” Seite 1
- photometric normalization
Textstelle · Seite 1
- “photometric normalization” Seite 1
- camera calibration
Textstelle · Seite 1
- “camera calibration” Seite 1
- pseudo-multi-view construction
Textstelle · Seite 1
- “pseudo-multi-view construction for foundation model adaptation” Seite 1
- lightweight finetuning
Textstelle · Seite 2
- “design lightweight finetuning, prompt tuning or LoRA-style specialization strategies” Seite 2
- prompt tuning
Textstelle · Seite 2
- “prompt tuning” Seite 2
- LoRA-style specialization
Textstelle · Seite 2
- “LoRA-style specialization strategies” Seite 2
- joint spatiotemporal learning head
Textstelle · Seite 2
- “Build a joint spatiotemporal learning head on top of the adapted 3D foundation backbone” Seite 2
- experimental baselines
Textstelle · Seite 2
- “Evaluate reconstruction quality (geometry fidelity, temporal consistency) and property inference accuracy against experimental baselines” Seite 2
- ablation studies
Textstelle · Seite 2
- “conduct ablation studies on foundation model adaptation strategies” Seite 2
- inference and training code
Textstelle · Seite 2
- “release reproducible inference and training code” Seite 2
- monocular video sequences
- Aufgaben
- Curate and preprocess monocular video sequences from rotating-drum granular flow experiments
Textstelle · Seite 1
- “Curate and preprocess monocular video sequences from rotating-drum granular flow experiments:” Seite 1
- Adapt existing pre-trained 3D foundation models to the granular flow domain
Textstelle · Seite 2
- “Adapt existing pre-trained 3D foundation models to the granular flow domain:” Seite 2
- Build a joint spatiotemporal learning head
Textstelle · Seite 2
- “Build a joint spatiotemporal learning head on top of the adapted 3D foundation backbone” Seite 2
- Evaluate reconstruction quality and property inference accuracy against experimental baselines
Textstelle · Seite 2
- “Evaluate reconstruction quality (geometry fidelity, temporal consistency) and property inference accuracy against experimental baselines” Seite 2
- Conduct ablation studies on foundation model adaptation strategies
Textstelle · Seite 2
- “conduct ablation studies on foundation model adaptation strategies” Seite 2
- Summarize results
Textstelle · Seite 2
- “Summarize results, document the adaptation pipeline, and release reproducible inference and training code for the 3D-aware granular world model.” Seite 2
- Document the adaptation pipeline
Textstelle · Seite 2
- “document the adaptation pipeline” Seite 2
- Release reproducible inference and training code
Textstelle · Seite 2
- “release reproducible inference and training code” Seite 2
- Curate and preprocess monocular video sequences from rotating-drum granular flow experiments
- Art der Arbeit
Voraussetzungen
- Erforderliche Kenntnisse
- Im PDF nicht angegeben
- Wünschenswerte Kenntnisse
- Strong proficiency in Python
Textstelle · Seite 2
- “Strong proficiency in Python;” Seite 2
- Solid computer vision background, with hands-on experience in video understanding, multi-view geometry and 3D reconstruction
Textstelle · Seite 2
- “Solid computer vision background, with hands-on experience in video understanding, multi-view geometry and 3D reconstruction;” Seite 2
- Proficient in PyTorch for deep learning research and model development
Textstelle · Seite 2
- “Proficient in PyTorch for deep learning research and model development;” Seite 2
- Practical experience with large-scale deep learning training: distributed training, dataset pipeline design, checkpoint management and hyperparameter tuning
Textstelle · Seite 2
- “Practical experience with large-scale deep learning training: distributed training, dataset pipeline design, checkpoint management and hyperparameter tuning;” Seite 2
- Familiarity with 3D / vision foundation models, including adaptation paradigms like LoRA, prompt tuning and domain alignment
Textstelle · Seite 2
- “Familiarity with 3D / vision foundation models (e.g. 3DGS, NeRF foundation models, point-cloud foundation models, vision-LiDAR foundation models), including adaptation paradigms like LoRA, prompt tuning and domain alignment;” Seite 2
- Curiosity for interdisciplinary research bridging 3D computer vision, foundation model adaptation and granular material physics
Textstelle · Seite 2
- “Curiosity for interdisciplinary research bridging 3D computer vision, foundation model adaptation and granular material physics.” Seite 2
- Strong proficiency in Python
- Geeignete Studienrichtungen
- Im PDF nicht angegeben
- Programmierung im Projekt
- Zentral
Textstelle · Seite 2
- “release reproducible inference and training code for the 3D-aware granular world model.” Seite 2
Rahmenbedingungen
- Teamgröße
- Im PDF nicht angegeben
- Dauer
- Im PDF nicht angegeben
- Beginn
- Im PDF nicht angegeben
- Bewerbungsfrist
- Im PDF nicht angegeben
- Arbeitsort
- 100% remote work possible
Textstelle · Seite 1
- “[100% remote work possible]” Seite 1
- Arbeitsmodus
- Remote
Textstelle · Seite 1
- “[100% remote work possible]” Seite 1
- Arbeitssprache
- Im PDF nicht angegeben
Lernen und Betreuung
- Lernmöglichkeiten
- End-to-end research experience covering video preprocessing, foundation model adaptation, temporal-consistent 3D reconstruction, and visual-to-physical property inference
Textstelle · Seite 2
- “You will gain end-to-end research experience covering video preprocessing, foundation model adaptation, temporal-consistent 3D reconstruction, and visual-to-physical property inference.” Seite 2
- End-to-end research experience covering video preprocessing, foundation model adaptation, temporal-consistent 3D reconstruction, and visual-to-physical property inference
- Angebotene Betreuung
- Im PDF nicht angegeben
Bewerbung und Kontakt
- Kontaktpersonen
- Yuan Tan · Yuan.tan@tum.de
Textstelle · Seite 2
- “Yuan Tan Gregor-Mendel-Straße 4, 85354 Freising Tel. +49 8161 71-5172 Yuan.tan@tum.de” Seite 2
- Yuan Tan · Yuan.tan@tum.de
- Bewerbung
- Im PDF nicht angegeben
- Weitere Informationen
- Im PDF nicht angegeben