Structuring Uncertainty for Fine-Grained Sampling in Stochastic Segmentation Networks

Authors: Frank Nussbaum, Jakob Gawlikowski, Julia Niebling

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show on medical-imaging, earth-observation, and traffic-scene data that rotation criteria based on factor-specific flow probabilities consistently yield the best factors for fine-grained sampling. 5 Experiments The purpose of our experiments is to (1) evaluate rotation criteria based on the quality of rotated factors, (2) demonstrate the merits of fine-grained sample control based on reasonably-rotated factors.
Researcher Affiliation Academia Frank Nussbaum Friedrich-Schiller-University Jena Fürstengraben 1, 07743 Jena, Germany and German Aerospace Center (DLR) Institute of Data Science Mälzerstraße 3-5, 07745 Jena, Germany frank.nussbaum@uni-jena.de Jakob Gawlikowski Technical University of Munich Arcisstraße 21, 80333 Munich, Germany and German Aerospace Center (DLR) Institute of Data Science Mälzerstraße 3-5, 07745 Jena, Germany jakob.gawlikowski@dlr.de Julia Niebling German Aerospace Center (DLR) Institute of Data Science Mälzerstraße 3-5, 07745 Jena, Germany julia.niebling@dlr.de
Pseudocode No The paper does not include a figure, block, or section labeled "Pseudocode" or "Algorithm", nor does it present structured steps in a code-like format.
Open Source Code Yes Additionally, we made the code for the proposed methods and experiments available under https://github.com/Jakob Code/Structuring SSNs.
Open Datasets Yes First, we use the LIDC data set [1] in its pre-processed version from [28]... Second, we use the multi-spectral Sentinel-2 data from the SEN12MS data set [36]... Third, we use the Cam Vid data set [5]...
Dataset Splits Yes Additional details and statistics about the data sets (including splits) can be found in the supplement, where we also detail all training procedures.
Hardware Specification Yes On a single core of an Intel Xeon Platinum 8260, factor-specific and full flow probabilities can be computed in the sub-second range without significant differences w.r.t. the used rotation, see the supplement for details.
Software Dependencies Yes We used Python 3.7, particularly with the libraries Py Torch 1.11 [32], scikit-learn [33], Num Py [18], and einops [34].
Experiment Setup Yes Additional details and statistics about the data sets (including splits) can be found in the supplement, where we also detail all training procedures.