Materials in Vessels Dataset, Annotated images of materials in transparent vessels for semantic segmentation

 Data set of materials in vessels
The handling of materials in glassware vessels is the main task in chemistry laboratory research as well as a large number of other activities. Visual recognition of the physical phase of the
materials is essential for many methods ranging from a simple task such as fill-level evaluation to the
identification of more complex properties such as solvation, precipitation, crystallization and phase
separation. To help train neural nets for this task, a new data set was created. The data set contains a
thousand images of materials, in different phases and involved in different chemical processes, in a
laboratory setting. Each pixel in each image is labeled according to several layers of classification, as
given below:

a. Vessel/Background: For each pixel assign value of one if it is part of the vessel and zero otherwise.
This annotation was used as the ROI map for the valve filter method.

b. Filled/Empty: This is similar to the above, but also distinguishes between the filled and empty
regions of the vessel. For each pixel, one of the following three values is assigned:0 (background); 1
(empty vessel); or 2 (filled vessel).

c. Phase type: This is similar to the above but distinguishes between liquid and solid regions of the
filled vessel. For each pixel, one of the following four values: 0 (background); 1 (empty vessel); 2
(liquid); or 3 (solid).

d. Fine-grained physical phase type: This is similar to the above but distinguishes between specific
classes of physical phase. For each pixel, one of 15 values is assigned: 1 (background); 2 (empty
vessel); 3 (liquid); 4 (liquid phase two, in the case where more than one phase of the liquid appears in
the vessel); 5 (suspension); 6 (emulsion); 7 (foam); 8 (solid); 9 (gel); 10 (powder); 11 (granular); 12
(bulk); 13 (solid-liquid mixture); 14 (solid phase two, in the case where more than one phase of solid
exists in the vessel): and 15 (vapor).
The annotations are given as images of the size of the original image, where the pixel value is the
class number. The annotation of the vessel region (a) is used in the ROI input for the valve filter net .

4.1. Validation/testing set
The data set is divided into training and testing sets. The testing set is itself divided into two subsets;
one contains images extracted from the same YouTube channels as the training set, and therefore was
taken under similar conditions as the training images. The second subset contains images extracted
from YouTube channels not included in the training set, and hence contains images taken under
different conditions from those used to train the net.

4.2. Creating the data set
The creation of a large number of images with a variety of chemical processes and settings could have
been a daunting task. Luckily, several YouTube channels dedicated to chemical experiments exist
which offer high-quality footage of chemistry experiments. Thanks to these channels, including
NurdRage, NileRed, ChemPlayer, it was possible to collect a large number of high-quality images in a
short time. Pixel-wise annotation of these images was another challenging task, and was performed by
Alexandra Emanuel and Mor Bismuth.

For more details see:  Setting attention region for convolutional neural  networks using region selective features, for  recognition of materials within glass vessels

This dataset was first published in 2017.8

For newer and Bigger datasets see

https://zenodo.org/record/4736111#.YbG-RrtyZH4

https://zenodo.org/record/3697452#.YbG-TLtyZH4