APPENDIX:

Abstract

We use t-SNE technique to embed representations of different human models into a 2D space (perplexity = 30 by default). Figure 1 shows a two-dimensional visualization of features on several models (three different humans, each person has twenty different poses). Each dot represents a different pose of a human model, with colour indicating human identity. We can see that shapes with different appearances are clearly separated, and shapes with the same appearances but in different poses are also separated as far as possible. Note that these representations are generated under five observed views.