Computer Vision News 6 Congrats, Doctor Yunlu! Yunlu Chen has recently completed his PhD at the University of Amsterdam, under the supervision of Efstratios Gavves, Thomas Mensink, and Arnold Smeulders. His research explored how 3D deep learning can effectively leverage the advantages of continuum and achieve better generalisation. Yunlu is now a postdoctoral researcher with Fernando De la Torre at Carnegie Mellon University. Continuity and discreteness are long-standing characteristics when juxtaposing natural and engineered systems. The natural world exhibits apparent analogue and continuous material and energy flows at our observable scale. In contrast, computing systems are based on the discrete nature of information processing, employing binary digit values and quantized domains. Traditional discrete 3D representations faces limitations in high memory and computation costs for high fidelity due to the need for elevated sampling rates as required by the Nyquist-Shannon theorem. To this end, the recent trend of modelling 3D signals with implicit neural representations is a new paradigm that utilizes continuous coordinatebased neural functions which are not bounded to any certain resolution. To understand how these representations encode 3D shapes, Yunlu’s paper at ICML 2021 investigated the hidden-layer features in latent-coded implicit representations, from which the emerging hierarchical structure is observed in the implicit network layers, such that the earlier layers encode coarse shape outlines, while deeper layers encode fine shape details (Fig. 1). In addition, this research suggests the representations’ unsupervised acquisition of correspondence and semantic awareness, which facilitates generalization to a collection of 3D shapes. Furthermore, Yunlu’s work at ECCV 2022 extended the representation by injecting the design of equivariance and graph embedding, allowing high-fidelity encoding of 3D signals in local details, and generalization to unseen geometric transformations including (continuous) rotation, translation and scaling (Fig. 2). Congrats, Doctor Yunlu!
RkJQdWJsaXNoZXIy NTc3NzU=