Hier finden Sie eine Auswahl der Poster der Abschlussveranstaltung:
Overall Methodology
KI-Absicherung: Overall Approach
Synthetic Data Generation based on a modern Game Engine
Physically based synthetic data generation pipeline
Pedestrian detector development using the SSD and the KI Absicherung synthetic dataset (synPeDS)
Data Structuring and Analysis
Synthetic Dataset for Pedestrian Detection (synPeDS) - Overview
Methodology of creating an ontology for dataset engineering
Motion Capture & Material Measurement
Lessons learned on Synthetic Data Generation
Method related Evidence Workstreams
Non-Parametric Uncertainty Optimization for Bounding Box Regression
Semantic Testing of DNNs with Proxy Models
Multivariate Confidence Calibration for Object Detection
Gradient Metrics for Uncertainty Quantification for Object Detection
Coverage Guided Fuzz Testing Framework & Dataset Quality Metrics
AugMix: Robustness via Data Augmentation
Visual Exploration and Semantic Analysis of DNN Weaknesses with ScrutinAI
Using concept embeddings for safety assurance
Data related Evidence Workstreams
Input Coverage Analysis using Domain Models and Combinatorial Testing
Parametrized, safety relevant test scenarios for DNN assessment
Applying Image Analysis, Combinatorial and Search-based Testing for DNN-Verification
Performance limiting factors analysis
Automated AI Validation using deep variational data synthesis
Methods & Measures
Morphological aggregation of heatmaps with Wasserstein k-means
Out-of-Distribution Detection in Semantic Segmentation
Extension of Deep Taylor Decomposition to Object Detection
Sensor Fusion for Robust Pedestrian Detection and Human Pose Estimation
Analysis and Comparison of Datasets by Leveraging Data Distributions in Latent Spaces
Self-compressing online pruning
Systematic Testing of DNNs Against Multiple Performance Limiting Factors (PLFs)
Safety & Testing
Perspectives on Safety: Estimating and Proving
Evidence-based safety argumentation: approach + organizational setup