Latent Space Visualization

About

Supported Browsers: Firefox 67+, Chrome 71+ Visualization of representation learned by deep learning model – U-Net trained on Inria Aerial Labeling Dataset - for building detection in satellite imagery task. Estimated performance measure Intersection over Union (IoU) is plotted as a background of scatter plot red color (IoU = 0) indicates that there was no overlap between detected and original object while deep blue color indicates good detection of a model (IoU = 1), it is meant to give possibility to qualitatively assess strength of predictions based on location in the latent space and its estimated score. The method presented gives a map of the latent landscape learned by the model.