Unsupervised Deep Shape Descriptor With Point Distribution Learning

Shi, Yi; Xu, Mengchen; Yuan, Shuaihang; Fang, Yi

doi:10.1109/cvpr42600.2020.00937

Cited by 17 publications

(29 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Unsupervised methods for learning shape descriptors follow two major lines of research, with the first line leveraging generative models such as autoencoders [Girdhar et al, 2016, Sharma et al, 2016, Yang et al, 2018 or generative adversarial networks (GANs) Wu et al [2016], Achlioptas et al [2018], Han et al [2019] and the second line focusing on probabilistic models [Xie et al, 2018, Shi et al, 2020. Autoencoder-based approaches focus either on adding additional supervision to the latent space via 2D predictability [Girdhar et al, 2016], adding de-noising [Sharma et al, 2016], or improving the decoder using a folding-inspired architecture [Yang et al, 2018].…”

Section: Related Workmentioning

confidence: 99%

“…GAN-based approaches leverage either an additional VAE structure [Wu et al, 2016], pre-training via earthmover or Chamfer distance [Achlioptas et al, 2018], or using inter-view prediction as a pretext task [Han et al, 2019]. For probabilistic methods, Xie et al [2018] proposes an energy-based convolutional network which is trained with Markov Chain Monte Carlo such as Langevin dynamics, and Shi et al [2020] proposes to model point clouds using a Gaussian distribution for each point. Of these approaches, only Shi et al [2020] focuses on producing robust representations.…”

Section: Related Workmentioning

confidence: 99%

“…For probabilistic methods, Xie et al [2018] proposes an energy-based convolutional network which is trained with Markov Chain Monte Carlo such as Langevin dynamics, and Shi et al [2020] proposes to model point clouds using a Gaussian distribution for each point. Of these approaches, only Shi et al [2020] focuses on producing robust representations.…”

Section: Related Workmentioning

confidence: 99%

“…The procedure for our shape classification experiment follows the established protocol for unsupervised shape classification evaluation: first, the network is pre-trained in an unsupervised manner using the ShapeNet dataset [Chang et al, 2015]. Using the embeddings from pre-training, either a 2-layer MLP [Shi et al, 2020] or linear SVM [Wu et al, 2015] is trained and evaluated on the ModelNet40 dataset. Following previous work [Wu et al, 2015, Shi et al, 2020, we only pre-train on the 7 major categories of ShapeNet (chairs, sofas, tables, boats, airplanes, rifles, and cars).…”

Section: Unsupervised Shape Classification Protocolmentioning

confidence: 99%

“…Previous unsupervised methods for learning shape descriptors have generally used either probabilistic models [Xie et al, 2018, Shi et al, 2020, generative adversarial networks (GANs) [Wu et al, 2015, Achlioptas et al, 2018, Han et al, 2019, or autoencoders [Girdhar et al, 2016, Sharma et al, 2016, Wu et al, 2015, Yang et al, 2018. One approach that has been relatively unexplored for deep learning methods but common in hand-crafted methods is to design shape descriptors that are invariant to transforms that preserve distances, either the extrinsic (Euclidean) distance [Belongie et al, 2001, Johnson and Hebert, 1999, Manay et al, 2004, Gelfand et al, 2005, Pauly et al, 2003 or intrinsic (geodesic) distance [Elad and Kimmel, 2003, Rustamov, 2007, Sun et al, 2009, Aubry et al, 2011.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Staying in Shape: Learning Invariant Shape Representations using Contrastive Learning

Gu,

Yeung

2021

Preprint

View full text Add to dashboard Cite

Creating representations of shapes that are invariant to isometric or almost-isometric transformations has long been an area of interest in shape analysis, since enforcing invariance allows the learning of more effective and robust shape representations. Most existing invariant shape representations are handcrafted, and previous work on learning shape representations do not focus on producing invariant representations. To solve the problem of learning unsupervised invariant shape representations, we use contrastive learning, which produces discriminative representations through learning invariance to user-specified data augmentations. To produce representations that are specifically isometry and almost-isometry invariant, we propose new data augmentations that randomly sample these transformations. We show experimentally that our method outperforms previous unsupervised learning approaches in both effectiveness and robustness.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Unsupervised Shape Classification Protocolmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Staying in Shape: Learning Invariant Shape Representations using Contrastive Learning

Gu,

Yeung

2021

Preprint

View full text Add to dashboard Cite

show abstract

A point cloud self-learning network based on contrastive learning for classification and segmentation

Zhou,

Wang,

Chen

et al. 2024

Vis Comput

View full text Add to dashboard Cite

Unsupervised Point Cloud Representation Learning by Clustering and Neural Rendering

Mei,

Saltori,

Ricci

et al. 2024

Int J Comput Vis

View full text Add to dashboard Cite

Data augmentation has contributed to the rapid advancement of unsupervised learning on 3D point clouds. However, we argue that data augmentation is not ideal, as it requires a careful application-dependent selection of the types of augmentations to be performed, thus potentially biasing the information learned by the network during self-training. Moreover, several unsupervised methods only focus on uni-modal information, thus potentially introducing challenges in the case of sparse and textureless point clouds. To address these issues, we propose an augmentation-free unsupervised approach for point clouds, named CluRender, to learn transferable point-level features by leveraging uni-modal information for soft clustering and cross-modal information for neural rendering. Soft clustering enables self-training through a pseudo-label prediction task, where the affiliation of points to their clusters is used as a proxy under the constraint that these pseudo-labels divide the point cloud into approximate equal partitions. This allows us to formulate a clustering loss to minimize the standard cross-entropy between pseudo and predicted labels. Neural rendering generates photorealistic renderings from various viewpoints to transfer photometric cues from 2D images to the features. The consistency between rendered and real images is then measured to form a fitting loss, combined with the cross-entropy loss to self-train networks. Experiments on downstream applications, including 3D object detection, semantic segmentation, classification, part segmentation, and few-shot learning, demonstrate the effectiveness of our framework in outperforming state-of-the-art techniques.

show abstract

Unsupervised Deep Shape Descriptor With Point Distribution Learning

Cited by 17 publications

References 31 publications

Staying in Shape: Learning Invariant Shape Representations using Contrastive Learning

Staying in Shape: Learning Invariant Shape Representations using Contrastive Learning

A point cloud self-learning network based on contrastive learning for classification and segmentation

Unsupervised Point Cloud Representation Learning by Clustering and Neural Rendering

Contact Info

Product

Resources

About