PCN: Point Completion Network

By Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, Martial Hebert. 2018

1,083 citations recorded by the source corpus.

Shape completion, the problem of estimating the complete geometry of objects from partial observations, lies at the core of many vision and robotics applications. In this work, we propose Point Completion Network (PCN), a novel learning-based approach for shape completion. Unlike existing shape completion methods, PCN directly operates on raw point clouds without any structural assumption (e.g. symmetry) or annotation (e.g. semantic class) about the underlying shape. It features a decoder design that enables the generation of fine-grained completions while maintaining a small number of parameters. Our experiments show that PCN produces dense, complete point clouds with realistic structures in the missing regions on inputs with various levels of incompleteness and noise, including cars from LiDAR scans in the KITTI dataset.

Subfields: Aerospace Engineering, Computational Mechanics, Computer Vision and Pattern Recognition, 3D Vision & Pose

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computational Mechanics

In Computer Vision and Pattern Recognition

In 3D Vision & Pose

Citations in the corpus

Multiview Compressive Coding for 3D Reconstruction (2023, methodology): “For objects, methods directly output full point clouds [21,79,80] or deploy generative models [76, 84], but are typically tied to a fixed resolution.”
DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation (2019, background): “shape completion networks of [16, 58] require preparing training data of partial shapes.”
Revealing Occlusions with 4D Neural Fields (2022, methodology): “In terms of predictions, we depict our non-ablated 4D dynamic scene completion model, the ablation without time (which essentially becomes a 3D scene completion model without dynamics), and ﬁnally the PCN [68] baseline.”
PU-GAN: A Point Cloud Upsampling Adversarial Network (2019, methodology): “Different from point completion [41, 10] which generates the entire object from partial input, point cloud upsampling tends to improve the point distribution uniformity within local patches.”
PoinTr: Diverse Point Cloud Completion with Geometry-Aware Transformers (2021, methodology): “Although the projection method proposed in [49] is a better approximation to real scans, our strategy is more flexible and efficient.”
Shape Prior Deformation for Categorical 6D Object Pose and Size Estimation (2020, methodology): “Specifically, we adopt the PointNet-like encoder proposed in [40], and a threelayer fully-connected decoder as shown in Fig.”
Deep Local Shapes: Learning Local SDF Priors for Detailed 3D Reconstruction (2020, methodology): “The methods use generative point cloud models for scene representation [3,57,58].”
Focal Sparse Convolutional Networks for 3D Object Detection (2022, methodology): “SIENet [23] builds upon PCN [56] for point completion in a twostage framework.”
Multimodal Virtual Point 3D Detection (2021, methodology): “In the future, we plan to apply learning-based methods [50, 70] to infer the detailed 3D shape and pose from both Lidar measurements and image features.”
Patch-Based Progressive 3D Point Set Upsampling (2018, methodology): “It has been successfully applied in point set segmentation [10, 40], generation [1,13,56], consolidation [14,45,58], deformation [57], completion [15, 60] and upsampling [58, 59, 61].”
Unsupervised Point Cloud Pre-training via Occlusion Completion (2020, methodology): “To show-case that our pre-training method is agnostic to architectures, we choose three different encoders, including Point-Net [37], PCN [60] and DGCNN [53].”
PF-Net: Point Fractal Network for 3D Point Cloud Completion (2020, methodology): “Later, Point Completion Network (PCN) [35] proposes the ﬁrst learning-based architecture focusing on shape completion task.”
TopNet: Structural Point Cloud Decoder (2019, methodology): “[38] The ground-truth in [38] was generated by uniformly sampling 16384 points on the mesh surfaces and the partial point clouds were generated by back-projecting 2.”
Unsupervised 3D Shape Completion through GAN Inversion (2021, background): “Since the pioneering work PCN [36], point cloudbased shape completion has seen significant development compared to other representation forms like meshes and voxel grids, due to its flexibility and popularity as a raw data format.”
GRNet: Gridding Residual Network for Dense Point Cloud Completion (2020, methodology): “Recent methods [9,11] adopt Multi-layer Perceptrons (MLPs) to directly operate on point clouds and aggregate information across points with max pooling.”
SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-Transformer (2021, methodology): “For evaluation, we adopt the L1 version of Chamfer distance, which follows the same practice as previous methods [57].”
Morphing and Sampling Network for Dense Point Cloud Completion (2019, methodology): “5% to that of AtlasNet and PCN, which demonstrates the superiority of our method.”