Support by the official sponsors of
the Single-Molecule Localization Microscopy Symposium

Single-Molecule Localization Microscopy  •  Software Benchmarking

Collection of reference datasets

The benchmarking of SMLM software package is based on the usage of common reference datasets. Here, we propose a collection of datasets under our conditions of use.

banner

Realistic Simulation

These datasets are simulated sequence of realistic frames. The bio-inspired samples are 3D continuous models that imitates biological structures. The frames are computed using advanced approximation of the PSF at the nanoscale and they include common perturbations of the sensor: different sources of noise, autofluorescence, etc.

picture1
Tubulins

This sample consists to a realistic structure of 7 tubulins (diameter 25 nm) and 1 tubulin (diameter 40 nm). 100000 fluorophores are activated over 2400 frames. Tubulins I has a low level of read-out noise autofluorescence background, and Tubulins II has the same structure with higher level of noise.

Tubulins I

Tubulins II

This dataset is a training dataset. The ground-truth localization and rendering image are only accessible only to registered teams or individuals.

Bundled of tubulins

This sample is a bundle of 8 tubes of 30 nm diameter. It exists in two versions: LS (long sequence) and HD (high density).

Long sequence

High density

This dataset is a training dataset. The ground-truth localization and rendering image are only accessible only to registered teams or individuals.

Contest Dataset 1

This sample consists to 8 tubulins of various diameters. The long-sequence (LS) dataset has 10'000 frames and the high-density dataset has 1'000 frames.

Long sequence

High density

This dataset is a contest dataset for evaluating the software The download information and frame are reserved to registered teams or individuals. The ground-truth localization will not be provided.

Contest Dataset 2

This sample is a dense network of tubulins. The long-sequence (LS) dataset has 1'200 frames and the high-density dataset has 204 frames.

Long sequence

High density

This dataset is a contest dataset for evaluating the software The download information and frame are reserved to registered teams or individuals. The ground-truth localization will not be provided.

Contest Dataset 3

This sample contains helicoidal tubes. It is very deep sample from 0 to 1μm. The long-sequence (LS) dataset has 7'000 frames and the high-density dataset has 600 frames. 7000 frames

Long sequence

High density

This dataset is a contest dataset for evaluating the software The download information and frame are reserved to registered teams or individuals. The ground-truth localization will not be provided.

Conditions of use

These reference datasets are designed to be largely used by the developpers to validate theirs software and by the users to check a software. They can be freely used if the sources and references are properly cited.

Some of these datasets have been used for the ISBI 2013 Grand Challenge organized in the conference IEEE ISBI 2013 at San Francisco to benchmark more than 25 software.

Real Experiments

The experimental datasets consist of a sequence of frames from real stained biologicial samples. The standard acquisition parameters are also provided.

picture1
Tubulin ConjAL647

Experimental sequence of 27'529 frames (128x128 pixels) with the parameters of acquisition.

Reference: Suliana Manley, Julia Gunzenhäuser and Nicolas Olivier, A starter kit for point-localization super-resolution imaging, Current Opinion in Chemical Biology 15, 2011.

Tubulins • Long Sequence

Experimental sequence of 1'500 frames with the parameters of acquisition
High density of fluorophores per frame.

Courtesy of Nicolas Olivier and Debora Keller, LEB, EPFL

picture1
Tubulins • High Density

Experimental sequence of 500 frames with the parameters of acquisition
High density of fluorophores per frame.

Courtesy of Nicolas Olivier and Debora Keller, LEB, EPFL

picture1
Tubulin AF647

The dataset represents a fixed cell, stained with mouse anti-alpha-tubulin primary antibody and Alexa647 secondary antibody. The intermittent increase in signal is due reactivation with a 405 nm laser.
This dataset is an experimental sequence of 9990 frames of 128x128 pixels.

Courtesy of Nicolas Olivier and Suliana Manley, LEB, EPFL

Artificial Datasets

These datasets represents artificial structure with some actived fluorophores at some specific postions. These datasets are mainly used to evaluated particular features of the localization software, like the ability to discriminate two neighboring fluorophores in the same frame.

picture1
Eye

This dataset consists to 4 non-overlapping tubes of 1nm of radius in a field of view 38.4 x 38.4 μm. Only 300 fluorophores are activated in 40 frames. It is an ideal dataset: no perturbation, no additional noise.

picture1
Snow

This dataset contains fluorophores are placed on a regular grid with various depth and various number of photons. The field of view is 50 μm x 50 μm

picture1
Seashell

The synthetic sample has been created articifically using a 3D artificial structure shape. This shape represents a seashell of six branches with various fluorophore densities. This sample is provided in two versions: flat sample (no Z) and thick sample (1 μm of depth).

© 2016 Biomedical Imaging Group, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
Last update: 19 Aug 2016