Benchmarking State-of-the-Art 3D Object Detectors for Logistics Applications Using Depth Cameras
- Institut
- Lehrstuhl für Fördertechnik Materialfluss Logistik
- Typ
- Semesterarbeit Masterarbeit
- Inhalt
- Beschreibung
Benchmarking State-of-the-Art 3D Object Detectors for Logistics Applications Using Depth Cameras
Objective:The objective of this thesis is to evaluate and compare different state-of-the-art 3D object detection algorithms concerning accuracy and inference speed using data from a depth camera. The focus is on assessing the practical applicability of these methods in logistics scenarios, where efficient and precise perception is essential.
Motivations:Accurate 3D object detection is essential in logistics to enable safe navigation, reliable automation, and efficient material handling. However, collecting and labeling large-scale 3D datasets in real warehouses is costly, time-consuming, and often impractical due to dynamic environments. Synthetic data offers a scalable and flexible alternative, allowing rapid dataset generation with precise ground truth. By training detection methods on synthetic data, it becomes possible to accelerate development, reduce costs, and adapt algorithms to logistics-specific scenarios that are difficult to capture in real-world settings.
Contributions:
The study will first investigate the availability of open-source implementations (e.g., GitHub repositories) and their support for dataset training. The final goal is to identify the best-performing method and train it on a synthetic logistics dataset, providing insights into its suitability for future robotic perception tasks in material handling and warehouse environments.
The selected algorithms include:
-
Frustum PointNets
-
RegionPLC
-
OV3D
-
Mosaic3D
-
Mask3D
-
SoftGroup
-
OneFormer3D
Each must be tested under the same conditions using our generated synthetic logistic dataset to ensure fair benchmarking.
-
- Voraussetzungen
Preconditions
-
Object detection & Machine Learning Knowledge – Understanding of depth sensing, point clouds, and modern 3D object detection methods.
-
Programming & Frameworks – Advanced Python skills and experience with PyTorch and GPU-accelerated training.
-
Data Handling & Preprocessing – Ability to work with depth camera data and synthetic datasets.
-
Benchmarking & Evaluation – Familiarity with 3D detection metrics, experiment design, and performance analysis.
-
Research & Reporting – Competence in literature review, adapting open-source code, and documenting results clearly.
-
- Möglicher Beginn
- sofort
- Kontakt
-
Daniel Vidal, M.Sc.
Raum: 5505.01.590C
Tel.: +49 (89) 289 - 15955
daniel.vidaltum.de - Ausschreibung
-
