Logan

Abstract

Fig. 1. We present LOGAN, a deep neural network which learns general-purpose shape transforms from unpaired domains. By altering only the two input data domains for training, without changing the network architecture or any hyper-parameters, LOGAN can transform between chairs and tables, from cross-sectional profiles to surfaces, as well as adding arms to chairs. It can also learn both style-preserving content transfer (letters R → P , A → H , in different font styles) and content-preserving style transfer (wide to narrow S , thick to thin I , thin to thick G, and italic to non-italic A.)We introduce LOGAN, a deep neural network aimed at learning generalpurpose shape transforms from unpaired domains. The network is trained on two sets of shapes, e.g., tables and chairs, while there is neither a pairing between shapes from the domains as supervision nor any point-wise correspondence between any shapes. Once trained, LOGAN takes a shape from one domain and transforms it into the other. Our network consists of an autoencoder to encode shapes from the two input domains into a common latent space, where the latent codes concatenate multi-scale shape features, resulting in an overcomplete representation. The translator is based on a generative adversarial network (GAN), operating in the latent space, where an adversarial loss enforces cross-domain translation while a feature preservation loss ensures that the right shape features are preserved for a natural shape transform. We conduct ablation studies to validate each of our key network designs and demonstrate superior capabilities in unpaired shape transforms on a variety of examples over baselines and state-of-the-art approaches. We show that LOGAN is able to learn what shape features to preserve during shape translation, either local or non-local, whether content or style, depending solely on the input domains for training.

Shape embedding and retrieval in multi-flow deformation

“…2.1 3D shape deformation 3D shape deformation aims to generate new shapes by deforming existing shapes while retaining local geometric features. Earlier works model deformation as an optimization problem to fit dense [16][17][18] or sparse key-point [19,20] observations with rigid [21] or non-rigid [22] regularization. Deformation can be free-form [23][24][25] where vertices [26][27][28] are directly optimized; alternatively, shape templates or cages [29][30][31][32] can serve as agents for deformation to preserve shape integrity.…”

USTNet: Unsupervised Shape‐to‐Shape Translation via Disentangled Representations

“…Yin et al [YHCZ18] pioneered the idea by developing a bidirectional point displacement network that learns geometric transformations between point sets from two domains. Yin et al proposed LOGAN [YCH*19], a general‐purpose deep neural network that learns shape‐to‐shape translation from unpaired inputs. LOGAN features an overcompleted autoencoder to explicitly assign features at different shape scales to different portions of the latent codes and a translator network to distinguish and translate the latent vector of the source shape to the target domain.…”

“…While such topics have been extensively explored in the past years, it is challenging to identify and transfer the style element for the modeling task, which plays a key role in conveying high‐level and abstract notions [HLvK*17]. Following this, and also related developments in stylized image processing [JYF*20], the style has gained increasing interest in the 3D modeling world in recent years [YGS*21], including the task of shape‐to‐shape translation [YCH*19,SGST20].…”

“…Several works have aimed to address the above goal. Yin et al proposed LOGAN [YCH*19], a general shape‐to‐shape translation framework that trains a translator to map a source latent code into a target latent code and achieves high‐quality cross‐domain shape transforms. Yet, the proposed translator can only generate a single translated result with one input shape.…”

Split-AE: An Autoencoder-based Disentanglement Framework for 3D Shape-to-shape Feature Transfer

“…The second task is to enforce the style space to learn in an unsupervised fashion the distinctive local features of 3D shapes from each shape class, such that it leads to an efficient disentanglement of class-essential and class-redundant information [12]. Further, following the idea of this unsupervised approach of shapeclass classifications to encode distinctive features in style space helps to extend our architecture to learn shapes from multiple classes, which is not possible with prior approaches [1], [13].…”