Hiwi position: 3D Scene Graph in OR Dataset and Benchmarking

Advisor: Evin Pinar Ornek
Starting date: End of July

Abstract

With the aim of holistic Operating Room modeling, we are creating a 4D dataset (point cloud videos) from a simulated surgical operation, with humans and different surgical equipment in the environment. To achieve this, the data should first be annotated with object bounding boxes, human poses and scene graph relationships. Then, different state-of-the-art approaches for each task should be trained to create a benchmark. To this end, your task would entail the further development of the labeling tools, using them for annotation, and initiating benchmarking efforts. There is a potential to continue in this project as a IDP/GR or master thesis.

Literature

[1] Johanna Wald, Helisa Dhamo, Nassir Navab, Federico Tombari, "Learning 3D Semantic Scene Graphs from 3D Indoor Reconstructions" CVPR 2020.

[2] Ege Ozsoy, Evin Pinar Ornek, Ulrich Eck, Federico Tombari, Nassir Navab, "Multimodal Semantic Scene Graphs for Holistic Modeling of Surgical Procedures", https://arxiv.org/abs/2106.15309, 2021.