Machine Learning

Autoencoder Art

In the examples below I experimented with sparse autoencoders to learn image features.

Green flowers.

Training image. Green flowers.

Hidden.

Features learned by 25 hidden units using 10000 randomly sampled 8×8 sub-images from the green flowers training image.

Test.

Test image (@ Copyright 2017 by Octavia Martin. Reproduced with permission).

Edges.

The feature encoding learned from the training image and applied to the test image on the left acts like an edge detector on the latter.

Vertical stripes.

Another training image. Vertical stripes (less “rich” in features than the green flowers).

Hidden.

Features learned by 25 hidden units using 10000 randomly sampled 8×8 sub-images from the vertical stripes training image.

Test.

Test image (@ Copyright 2017 by Octavia Martin. Reproduced with permission).

Edges.

The feature detector learned from the vertical stripes image and applied to the same test image recognizes fewer features (i.e., only vertical edges).

 

Clustering

Clustering can be used as a form of unsupervised learning of geometric features suitable for re-parameterizing models for subsequent texture and / or geometry editing. In this first example, flat-ish regions on 3D models are learned for the purpose of 2D texture painting. See our ACM Siggraph sketch for details.

Siggraph sketch

As another example, quadrilateral parameterization domains for arbitrary meshes are derived via multi-layer normal clustering and used to generate multi-resolution subdivision surfaces suitable for interactive editing.

Quad domain parameterization

Additional information in Domain Decomposition for Multiresolution Analysis, Symposium on Geometry Processing 2003, Aachen, Germany.