4.7 Results: Evaluating LSBD withDLSBD 91 observe some consistency in object identity, see Figure 4.11. Such data generation cannot directly be obtained from traditional disentanglement methods since there is no clear direction representing either the object identities or the orientations. Object interpolation Next, we show how the latent space is structured among the latent variables representing the object identities for models trained on COIL-100. We show the generated data obtained from decoding linearly interpolated latent variables between different objects to show the transitions between objects and orientations. (a) Interpolation across Z =S1×RD. (b) Interpolation across Z =RD. Figure 4.12: Diagrams illustrating the interpolation between the latent variables associated to two objects. (a) Interpolation across a hyper-cylinder within Z =S1 ×RD used by LSBD-VAE. (b) Interpolation across Z =RD of traditional disentanglement models. In the traditional disentanglement models the linear interpolation can show the crossing of the latent codes associated to unexpected objects. For LSBD-VAE the interpolation is simple, as visualised in Figure 4.12a. First, we estimate the latent variables associated to the identities of the start and end objects by averaging the Euclidean latent variables of all images per object. Second, we calculate the linear interpolation between the object identity latent variables of the start and end object to generate a path through the object identity space. Finally, we combine the estimated identity variables in the path with
RkJQdWJsaXNoZXIy MjY0ODMw