110 Out-of-Distribution Generalisation with LSBD Representations (2021) show that disentanglement methods trained on data with correlations between the factors of variation learn representations that reflect these correlations, which further emphasises the need for better disentanglement methods that can generalise to OOD data. 5.3 Experimental Setup We investigate the out-of-distribution (OOD) generalisation of a number of VAEbased disentanglement models for datasets with known factorised generative factors, by splitting off data with certain factor combinations into an OOD set and using the remaining data as training set. We then evaluate those models on their performance on the left-out OOD data using various metrics in Section 5.4. 5.3.1 Datasets We consider two datasets with an underlying SO(2) ×SO(2) group structure (i.e. having 2 cyclic factors), Square and Arrow; as well as two popular disentanglement datasets, dSprites and 3D Shapes. See Figure 5.2 for some examples. The datasets can be fully generated by known factors of variation. All datasets contain images with64×64 pixels (black-and-white or RGB). Figure 5.2: Example images of Square, Arrow, dSprites, and 3D Shapes. Square has factors x-position andy-position, squares wrap around the edges of the canvas so the factors are cyclic. Arrow has factors orientation and hue.
RkJQdWJsaXNoZXIy MjY0ODMw