This paper proposes a fully differentiable and end-to-end framework for learning Bézier decomposition on 3D point clouds. The framework aims to partition input point clouds into multiple Bézier primitive patches through a learned Bézier decomposition process. Unlike previous approaches that handle different primitive types separately, thus being limited to specific shape categories, our method seeks to achieve a generalized primitive segmentation on point clouds. Drawing inspiration from Bézier decomposition on NURBS models, we adapt it to guide point cloud segmentation without relying on pre-defined primitive types. To achieve this, we introduce a joint optimization framework that simultaneously learns Bézier primitive segmentation and geometric fitting in a cascaded architecture. Additionally, we propose a soft voting regularizer to enhance primitive segmentation and an auto-weight embedding module to effectively cluster point features, making the network more robust and applicable to various scenarios. Furthermore, we incorporate a reconstruction module capable of processing multiple CAD models with different primitives simultaneously. Extensive experiments were conducted on both synthetic ABC datasets and real-scan datasets to validate and compare our approach against several baseline methods. The results demonstrate that our method outperforms previous work in terms of segmentation accuracy, while also exhibiting significantly faster inference speed.

With recent success of deep learning in 2-D visual recognition, deep-learning-based 3-D point cloud analysis has received increasing attention from the community, especially due to the rapid development of autonomous driving technologies. However, most existing methods directly learn point features in the spatial domain, leaving the local structures in the spectral domain poorly investigated. In this article, we introduce a new method, PointWavelet, to explore local graphs in the spectral domain via a learnable graph wavelet transform. Specifically, we first introduce the graph wavelet transform to form multiscale spectral graph convolution to learn effective local structural representations. To avoid the time-consuming spectral decomposition, we then devise a learnable graph wavelet transform, which significantly accelerates the overall training process. Extensive experiments on four popular point cloud datasets, ModelNet40, ScanObjectNN, ShapeNet-Part, and S3DIS, demonstrate the effectiveness of the proposed method on point cloud classification and segmentation.

This paper aims to predict accurate unoriented normals from unorganized 3D point clouds, possibly containing noise and varying densities. Previous methods for unoriented normal estimation either fit an explicit surface from a local patch to infer normals or directly regress normals using neural networks. However, both approaches have limitations. Fitting explicit surfaces can be susceptible to challenging regions like sharp features due to the limited representation power of explicit surfaces. Meanwhile, directly regressing normals using neural networks lacks geometric interpretation. Different from previous works, we propose estimating normals by learning a neural implicit surface for a local patch. Our proposed neural implicit surface function has greater representation flexibility, increasing accuracy and robustness, even for complex shapes. Furthermore, we incorporate positional encoding with self-attention into our neural implicit surface function, refining the estimated normals further. Experimental results demonstrate that the proposed approach outperforms existing baselines on the PCPNet and SceneNN datasets, achieving state-of-the-art results.

Reference line detection is a challenging problem due to localization uncertainty and severe occlusion. To deal with the two issues, we propose a general framework for reference line detection with two modules: Gaussian line detection and connection graph inference. The first module outputs a set of Gaussian blurred lines, outlining the main compositions of the input image. For lines obscured by occlusions, the second module generates a connection graph of detected key point pairs by equidistant sampling on the feature map. Experimental results show that the proposed method could extract reference lines accurately and reliably in different application scenarios.

Neural Radiance Field (NeRF) stands out by demonstrating photo-realistic renderings, while it suffers from quality degradation when given only a few shot inputs. This paper aims to improve the rendering quality of NeRF from a few shot inputs. Original NeRF tends to overfit input views rapidly at the initial training stage when trained on sparse inputs. We address this challenge by presenting a novel incremental tensor decomposition method, where the resolution of decomposed tensors increases with the training iteration, enabling coarse to fine learning and alleviating the overfitting during the early stage. We offer several regularizations based on the proposed incremental learning process, including patchbased density regularization and depth regularization. Our proposed method outperforms previous baselines on the Realistic Synthetic 360° dataset and achieves state-of-the-art results in PSNR.

This paper addresses the challenge of accurately estimating unoriented normals for 3D point clouds, particularly in the presence of noise, varying densities, and sharp features. To overcome these challenges, we introduce a novel network architecture that integrates a geometry-weighted transformer module with a dual backbone network. This combined architecture enables accurate prediction of normals from patch embeddings that encode edge and spatial features. Additionally, we propose a region-growing loss to supervise the network, which can aid in normal estimation for sharp features by segmenting input patches based on normal discontinuities, thus avoiding the over-smoothness for normals in the sharp feature areas. Extensive experiments on PCPNet and SceneNN datasets highlight the effectiveness of our approach, leading to state-of-the-art performance against baseline methods.

Neural Radiance Field (NeRF) has recently demonstrated impressive photo-realistic renderings when trained on dense input views. However, it suffers from performance degradation from sparse inputs due to overfitting and inaccurate scene geometry estimation. This work addresses those challenges in few-shot neural radiance fields. Specifically, we propose initializing the training with pseudo-views augmented by forward warping using sparse inputs and Structure from Motion (COLMAP) obtained depth to prevent over-fitting at the start of training. Then, we take inspiration from classical image-based rendering techniques, and aggregate and blend features extracted from training views to supervise the distribution of rendered novel views. Furthermore, our method introduces a novel regularization by modeling the distribution of color and density. Our proposed approach outperforms baselines in the standard benchmark (e.g. DTU dataset [1]), achieving state-of-the-art performance.

Point cloud sampling is crucial for efficient large-scale point cloud analysis, where learning-to-sample methods have recently received increasing attention from the community for jointly training with downstream tasks. However, the abovementioned task-specific sampling methods usually fail to explore the geometries of objects in an explicit manner. In this paper, we introduce a new skeleton-aware learning-to-sample method by learning object skeletons as the prior knowledge to preserve the object geometry and topology information during sampling. Specifically, without labor-intensive annotations per object category, we first learn category-agnostic object skeletons via the medial axis transform definition in an unsupervised manner. With object skeleton, we then evaluate the histogram of the local feature size as the prior knowledge to formulate skeleton-aware sampling from a probabilistic perspective. Additionally, the proposed skeleton-aware sampling pipeline with the task network is thus end-to-end trainable by exploring the reparameterization trick. Extensive experiments on three popular downstream tasks, point cloud classification, retrieval, and reconstruction, demonstrate the effectiveness of the proposed method for efficient point cloud analysis.

This paper proposes BPNet, a novel end-to-end deep learning framework to learn Bézier primitive segmentation on 3D point clouds. The existing works treat different primitive types separately, thus limiting them to finite shape categories. To address this issue, we seek a generalized primitive segmentation on point clouds. Taking inspiration from Bézier decomposition on NURBS models, we transfer it to guide point cloud segmentation casting off primitive types. A joint optimization framework is proposed to learn Bézier primitive segmentation and geometric fitting simultaneously on a cascaded architecture. Specifically, we introduce a soft voting regularizer to improve primitive segmentation and propose an auto-weight embedding module to cluster point features, making the network more robust and generic. We also introduce a reconstruction module where we successfully process multiple CAD models with different primitives simultaneously. We conducted extensive experiments on the synthetic ABC dataset and real-scan datasets to validate and compare our approach with different baseline methods. Experiments show superior performance over previous work in terms of segmentation, with a substantially faster inference speed.

Generative models for 3D point clouds are extremely important for scene/object reconstruction applications in autonomous driving and robotics. Despite recent success of deep learning-based representation learning, it remains a great challenge for deep neural networks to synthesize or reconstruct high-fidelity point clouds, because of the difficulties in 1) learning effective pointwise representations; and 2) generating realistic point clouds from complex distributions. In this paper, we devise a dual-generators framework for point cloud generation, which generalizes vanilla generative adversarial learning framework in a progressive manner. Specifically, the first generator aims to learn effective point embeddings in a breadth-first manner, while the second generator is used to refine the generated point cloud based on a depth-first point embedding to generate a robust and uniform point cloud. The proposed dual-generators framework thus is able to progressively learn effective point embeddings for accurate point cloud generation. Experimental results on a variety of object categories from the most popular point cloud generation dataset, ShapeNet, demonstrate the state-of-the-art performance of the proposed method for accurate point cloud generation.