The new University of Duisburg-Essen diatom digital image training data set (UDE did it v1.0) - Université de Lorraine Access content directly
Poster Communications Year : 2023

The new University of Duisburg-Essen diatom digital image training data set (UDE did it v1.0)

Abstract

Here, we introduce a collection of more than 80,000 light microscopy images of individual diatom valves or frustules, covering a broad range of taxa and morphology by more than 300 samples from 15 different river and lake ecotypes. The diatoms were imaged at high resolution (< 0.1 µm/pixel) by transmitted light bright-field microscopy, with focus stacking to artificially increase focal depth up to 25 µm, allowing simultaneous observation of valve ornamentation and shape. Taken from a real-world setting, the images partly also include debris, mineral particles or other diatoms. Four experts identified over 500 diatom species; more than 100 species are represented by at least 100 specimens and about 150 by at least 50 specimens each. This data set is about one order of magnitude larger than previously published diatom data sets, and its high interspecies similarity makes it a valuable resource e.g. for benchmarking fine-grained out-of-distribution (OOD) detection, on which we present preliminary results.
No file

Dates and versions

hal-04187561 , version 1 (24-08-2023)

Identifiers

  • HAL Id : hal-04187561 , version 1

Cite

Michael Kloster, Andrea Burfeid-Castellanos, Danijela Vidacović, Ntambwe Albert Serge Mayombo, Mimoza Dani, et al.. The new University of Duisburg-Essen diatom digital image training data set (UDE did it v1.0). 16th International Diatom Symposium, Aug 2023, Yamagata, Japan. ⟨hal-04187561⟩
37 View
0 Download

Share

Gmail Mastodon Facebook X LinkedIn More