HI! I'm Robin Baumann
Computer Vision and Deep Learning Researcher
from Karlsruhe, Germany.

Robin Baumann, Jan-Friso Evers-Senne, Marcus Strand
Published in: 2019 IEEE International Conference on Industrial Cyber Physical Systems (ICPS)

The objective of this paper is to propose an approach which combines the multi view 3D reconstruction process of a building's facade with semantic information obtained by a deep neural network on the 2D observations of the considered scene. The object detector was used to annotate precise 2D features with semantic labels based on the object proposals. In a next step, spatially connected 2D features were evaluated to extend the initially computed labels to the whole facade element. The set of annotated 2D precise features was used to derive a classification of reconstructed 3D features. The a priori probabilities of the object detection model were used to compensate for biased predictions. The described approach was evaluated thoroughly with a view to different strategies for the annotation correction in 2D. Several failure patterns occurring in the considered application of facade reconstruction were observed and suggested solutions were given.

Robin Baumann, David Böhmb, Raoul Saiptb and Astrid Laubenheimer
Presentes at the 8th Collaborative European Research Conference (CERC) June 9-10, 2023, Barcelona, Spain, 2023. Read the PDF.

Homomorphic encryption (HE) enables privacy-preserving deep learning, but it comes with significant perfor- mance overheads. In this study, we evaluate the impact of model architectures on the utility and efficiency of deep learning models under differential privacy (DP) and HE settings. Our experiments reveal that dedicated model architectures are crucial for maintaining model utility when using DP. Moreover, we observe that align- ing complex model architectures like ResNets for HE by replacing ReLU with square activation, max pooling with average pooling, and group norm with batch norm strongly deteriorates model utility and results in archi- tectures with sharp minima that fail to generalize. Training such models with DP, however, yields a regularizing effect that improves model utility. Our study contributes to the understanding of the role of model architecture on the applicability of DP and HE.