Click on the below button to generate a random picture with geometric shapes, and to display the corresponding label predictions.

Raw image

Ground-truth

Prediction

Training procedure

The model used here has a simple architecture where three transposed
convolution layers follow three standard convolution layers. It has been
trained during 30 epochs with a set of 18000 randomly generated images
involving an other set of 2000 random images at the end of each training
epoch..

Description

This toy dataset contains three labels, i.e. squares, circles
and triangles. There is at most one single items for each shape: as a
consequence, some pictures may be empty!