Diffusion Probabilistic Models for 3D Point Cloud Generation

Luo, Shitong; Hu, Wei

doi:10.1109/cvpr46437.2021.00286

Cited by 258 publications

(157 citation statements)

References 7 publications

Supporting

Mentioning

126

Contrasting

Order By: Relevance

“…Even though only two SPDs are used in the decoder, our method still achieves the best performance among the compared counterparts in terms of both L 2 CD and EMD. Especially on the whole ShapeNet dataset, SnowflakeNet reduces the average CD by 1.05, which is 20% lower than DPM [76] (5.25 in terms of average CD). At the same time, our method also outperforms the other methods across all categories in terms of per-category CD and EMD.…”

Section: Quantitative Comparisonmentioning

confidence: 81%

“…To fairly evaluate the generation ability of SPD, we follow the same experimental settings of DPM [76] and conduct point cloud auto-encoding on the ShapeNet [70] dataset. The ShapeNet [70] dataset contains 51127 shapes from 55 categories, we use the same training, testing and evaluation split of DPM [76], where the ratio of training, testing and validation sets are 80%, 15% and 5%, respectively. Four datasets are used in evaluation, which include three categories (i.e.…”

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

“…Dataset. For the task of point cloud generation, we follow DPM [76] and conduct experiments on the ShapeNet dataset. We quantitatively compare our method with several state-of-the-art generative models on the two categories (i.e.…”

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

“…Evaluation metric. To evaluate the generation quality of our method, we follow prior works [74], [75], [76] and employ the coverage score (COV), the minimum matching distance (MMD), 1-NN classifier accuracy (1-NNA) and the Jenson-Shannon divergence (JSD). The COV score measures whether the generated samples cover all the modes of the data distribution and the MMD score measures the fidelity of the generated samples.…”

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

See 3 more Smart Citations

Snowflake Point Deconvolution for Point Cloud Completion and Generation with Skip-Transformer

Peng¹,

Xu²,

Liu³

et al. 2022

Preprint

View full text Add to dashboard Cite

Most existing point cloud completion methods suffered from discrete nature of point clouds and unstructured prediction of points in local regions, which makes it hard to reveal fine local geometric details. To resolve this issue, we propose SnowflakeNet with Snowflake Point Deconvolution (SPD) to generate the complete point clouds. SPD models the generation of complete point clouds as the snowflake-like growth of points, where the child points are progressively generated by splitting their parent points after each SPD. Our insight of revealing detailed geometry is to introduce skip-transformer in SPD to learn point splitting patterns which can fit local regions the best. Skip-transformer leverages attention mechanism to summarize the splitting patterns used in previous SPD layer to produce the splitting in current SPD layer. The locally compact and structured point clouds generated by SPD precisely reveal the structure characteristic of 3D shape in local patches, which enables us to predict highly detailed geometries. Moreover, since SPD is a general operation, which is not limited to completion, we further explore the applications of SPD on other generative tasks, including point cloud auto-encoding, generation, single image reconstruction and upsampling. Our experimental results outperform the state-of-the-art methods under widely used benchmarks.

show abstract

Section: Quantitative Comparisonmentioning

confidence: 81%

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

Section: Dataset and Evaluation Metricmentioning

confidence: 99%

See 2 more Smart Citations

Snowflake Point Deconvolution for Point Cloud Completion and Generation with Skip-Transformer

Peng¹,

Xu²,

Liu³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Denoising Diffusion Probabilistic Models (DDPM) (Sohl-Dickstein et al, 2015;Song & Ermon, 2019;Ho et al, 2020) have emerged as a powerful family of generative models, capable of synthesizing high-quality images, audio, and 3D shapes (Ho et al, 2020;Chen et al, 2021a;b;Cai et al, 2020;Luo & Hu, 2021). Recent work shows that DDPMs can outperform Generative Adversarial Networks (GAN) (Goodfellow et al, 2014;Brock et al, 2018) in generation quality, but unlike GANs, DDPMs admit likelihood computation and much more stable training dynamics Gulrajani et al, 2017).…”

Section: Introductionmentioning

confidence: 99%

Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality

Watson¹,

Chan²,

Ho³

et al. 2022

Preprint

View full text Add to dashboard Cite

Diffusion models have emerged as an expressive family of generative models rivaling GANs in sample quality and autoregressive models in likelihood scores. Standard diffusion models typically require hundreds of forward passes through the model to generate a single high-fidelity sample. We introduce Differentiable Diffusion Sampler Search (DDSS): a method that optimizes fast samplers for any pre-trained diffusion model by differentiating through sample quality scores. We present Generalized Gaussian Diffusion Models (GGDM), a family of flexible non-Markovian samplers for diffusion models. We show that optimizing the degrees of freedom of GGDM samplers by maximizing sample quality scores via gradient descent leads to improved sample quality. Our optimization procedure backpropagates through the sampling process using the reparametrization trick and gradient rematerialization. DDSS achieves strong results on unconditional image generation across various datasets (e.g., FID scores on LSUN church 128x128 of 11.6 with only 10 inference steps, and 4.82 with 20 steps, compared to 51.1 and 14.9 with strongest DDPM/DDIM baselines). Our method is compatible with any pre-trained diffusion model without fine-tuning or re-training required.

show abstract

Point cloud synthesis with stochastic differential equations

Wang

Liu

et al. 2023

Computer Animation & Virtual

View full text Add to dashboard Cite

In this article, we propose a point cloud synthesis method based on stochastic differential equations. We view the point cloud generation process as smoothly transforming from a known prior distribution toward the high‐likelihood shape by point‐level denoising. We introduce a conditional corrector sampler to improve the quality of point clouds. By leveraging Markov chain Monte Carlo sample, our method can synthesize realistic point clouds. We additionally prove that our approach can be trained in an auto‐encoding fashion and reconstruct the point cloud faithfully. Furthermore, our model can be extended on a downstream application of point cloud completion. Experimental results demonstrate the effectiveness and efficiency of our method.

show abstract

Diffusion Probabilistic Models for 3D Point Cloud Generation

Cited by 258 publications

References 7 publications

Snowflake Point Deconvolution for Point Cloud Completion and Generation with Skip-Transformer

Snowflake Point Deconvolution for Point Cloud Completion and Generation with Skip-Transformer

Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality

Point cloud synthesis with stochastic differential equations

Contact Info

Product

Resources

About