IPAB Workshop-04/06/2020 Ian Mason Title: Modelling Locomotion Styles for Humanoid Animation with Feature-wise Transformations Abstract: I will be discussing ongoing work on varying the style of locomotion performed by a 3D character using a neural network animation system. We investigate the application of feature-wise transformations in the hidden layers of the network to allow for modifying style in real-time and demonstrate that this method can create reasonable interpolations between seen styles. I will also talk briefly about future directions for this work. Nanbo Li Title: Learning Neural Scene Representations of Multi-object Scenes With Multi-view Observations Abstract: Object-centric scene representations have the potential to support high-level cognitive abilities like causal reasoning and object-centric exploration. Current approaches for generative scene representation learning research are either inaccurate (suffering from single-view spatial ambiguity) or unstructured (failing to attain object-level scene factorization and interpretation). In this work, we address multi-view and multi-object representation and inference to resolve spatial ambiguity with spatial exploration. Through experiments we demonstrate that our method not only learns more accurate and disentangled object-centric representations, but also does it allow prediction of both observations and object segmentation for novel viewpoints. Jun 04 2020 13.00 - 14.00 IPAB Workshop-04/06/2020 Nanbo Li, Ian Mason, Kuba Sanak Blackboard Collaborate https://eu.bbcollab.com/guest/7d9a6997dfb04dce850edf1fa679f608
IPAB Workshop-04/06/2020 Ian Mason Title: Modelling Locomotion Styles for Humanoid Animation with Feature-wise Transformations Abstract: I will be discussing ongoing work on varying the style of locomotion performed by a 3D character using a neural network animation system. We investigate the application of feature-wise transformations in the hidden layers of the network to allow for modifying style in real-time and demonstrate that this method can create reasonable interpolations between seen styles. I will also talk briefly about future directions for this work. Nanbo Li Title: Learning Neural Scene Representations of Multi-object Scenes With Multi-view Observations Abstract: Object-centric scene representations have the potential to support high-level cognitive abilities like causal reasoning and object-centric exploration. Current approaches for generative scene representation learning research are either inaccurate (suffering from single-view spatial ambiguity) or unstructured (failing to attain object-level scene factorization and interpretation). In this work, we address multi-view and multi-object representation and inference to resolve spatial ambiguity with spatial exploration. Through experiments we demonstrate that our method not only learns more accurate and disentangled object-centric representations, but also does it allow prediction of both observations and object segmentation for novel viewpoints. Jun 04 2020 13.00 - 14.00 IPAB Workshop-04/06/2020 Nanbo Li, Ian Mason, Kuba Sanak Blackboard Collaborate https://eu.bbcollab.com/guest/7d9a6997dfb04dce850edf1fa679f608
