Classify on the Clock (CloCk) - An Entry Level Image Data Set for Neural Networks

General information

This data set contains synthetic images of an analog clock. Each time point from 00:00:00 - 11:59:59 is included as a separate image. We provide three different versions of each image with an increasing number of additional information:

Sparse: The image includes just the hour, minute and second hand

Reduced: Additional markings around the clock

Full: Additional written numbers

The hands differ in size and color:

Hour: Black, short, wide

Minute: Blue, long, medium

Second: Red, long, slim

We provide the following files in this data repository:

RGB images with a size of 512x512 for all three versions

A suggested training, validation and test split

The creation script as Python file

The script can easily be modified to create images with a different size and color.

Clock System

The movements of the hands follow a linear relationship. Their angles can be calculated by the forward system:

\(
\begin{bmatrix}
6^\circ & 0 & 0 \\
0.1^\circ & 6^\circ & 0 \\
0 & 0.5^\circ & 30^\circ
\end{bmatrix}
\begin{pmatrix}
n_{\text{sec}} \\
n_{\text{min}} \\
n_{\text{hour}}
\end{pmatrix}

=
\begin{pmatrix}
\alpha_{\text{sec}}\\
\alpha_{\text{min}} \\
\alpha_{\text{hour}}
\end{pmatrix}.\)

One can also introduce rotated versions of the images. The system becomes non-linear in this case:

\(\operatorname{mod}\left(
\begin{bmatrix}
6^\circ & 0 & 0 \\
0.1^\circ & 6^\circ & 0 \\
0 & 0.5^\circ & 30^\circ
\end{bmatrix}
\begin{pmatrix}
n_{\text{sec}} \\
n_{\text{min}} \\
n_{\text{hour}}
\end{pmatrix}
+ \omega, \, 360^\circ
\right)
=
\begin{pmatrix}
\alpha_{\text{sec}}\\
\alpha_{\text{min}} \\
\alpha_{\text{hour}}
\end{pmatrix}.\)

Use Cases

This data set was originally designed for basic research on Capsule Networks. Use cases are:

Evalutation of classification and regression performance

Influence of image transformations, e.g. rotations

Detection of the hierarchy & relationship of image parts

Concealment of objects in the image

Interpretability of the learned model

Solving a discrete inverse problem (sparse version)

...