Abstract:Point cloud completion aims to infer a complete shape from its partial observation. Many approaches utilize a pure encoderdecoder paradigm in which complete shape can be directly predicted by shape priors learned from partial scans, however, these methods suffer from the loss of details inevitably due to the feature abstraction issues. In this paper, we propose a novel framework,termed SPAC-Net, that aims to rethink the completion task under the guidance of a new structural prior, we call it interface. Specifically, our method first investigates Marginal Detector (MAD) module to localize the interface, defined as the intersection between the known observation and the missing parts. Based on the interface, our method predicts the coarse shape by learning the displacement from the points in interface move to their corresponding position in missing parts. Furthermore, we devise an additional Structure Supplement(SSP) module before the upsampling stage to enhance the structural details of the coarse shape, enabling the upsampling module to focus more on the upsampling task. Extensive experiments have been conducted on several challenging benchmarks, and the results demonstrate that our method outperforms existing state-of-the-art approaches.

What problem does this paper attempt to address?

This paper attempts to solve the problem of detail loss in point cloud data in the 3D object shape completion task. Specifically, many existing methods rely on the pure encoder - decoder paradigm to directly predict the complete shape by learning shape priors from partial scans. However, these methods will inevitably lose details due to the feature abstraction problem. ### Main problems 1. **Detail loss**: When performing point cloud completion, existing methods often lead to the loss of detailed information because of feature abstraction (such as the max - pooling operation). 2. **Insufficient geometric priors**: Many methods lack geometric sensitivity to the missing parts, resulting in the generated shapes being inaccurate in terms of boundaries and details. 3. **Problems of simultaneous up - sampling and detail enhancement**: Some methods perform up - sampling and detail enhancement simultaneously in the same stage, which may introduce noise and affect the quality of the generated results. ### Solutions To solve the above problems, the author proposes a new framework named SPAC - Net, and the main contributions include: 1. **Introducing a new structural prior - Interface**: - Define the interface as the intersection between the known observed part and the missing part. - By locating the interface, guide the model to perceive and complete the missing part more accurately, avoiding being affected by other noise information. 2. **Learning displacements instead of directly generating rough shapes**: - By learning the displacements from the interface points to their corresponding missing part positions, instead of directly generating rough shapes based on shape priors, effectively alleviates the problem of local feature loss. 3. **Introducing the MAD module and the SSP module**: - **MAD module**: Responsible for locating the interface. - **SSP module**: Enhances the structural details of the rough shape before shape up - sampling, supports the up - sampling module to obtain high - quality input, and further restores the shape details. 4. **Completing the task in stages**: - Divide the completion task into two main stages: shape generation and shape up - sampling. - In the shape generation stage, generate a rough shape by learning the displacements from the interface points to the missing part. - In the shape up - sampling stage, use the SSP module to optimize the structural details of the rough shape, and then perform high - resolution restoration of the missing part. ### Experimental results The author has carried out extensive experiments on multiple challenging benchmark datasets, and the results show that SPAC - Net significantly outperforms existing methods in the point cloud completion task, especially in maintaining structural characteristics and uniform distribution. Through these improvements, SPAC - Net not only improves the accuracy of point cloud completion but also enhances the sensitivity to the details of the missing part, solving the problems of detail loss and insufficient geometric priors in existing methods.

SPAC-Net: Rethinking Point Cloud Completion with Structural Prior

Point Cloud Completion Cascade Optimization Network Based on Feature Fusion

Adaptive Recurrent Forward Network for Dense Point Cloud Completion

Dual-scale Point Cloud Completion Network Based on High-Frequency Feature Fusion

RFNet - Recurrent Forward Network for Dense Point Cloud Completion.

Up-to-Down Network: Fusing Multi-Scale Context for 3D Semantic Scene Completion

Point Projection Network: A Multi-View-Based Point Completion Network with Encoder-Decoder Architecture

Point Cloud Completion by Skip-attention Network with Hierarchical Folding

Point Cloud Completion via Multi-Scale Edge Convolution and Attention

Point cloud completion via structured feature maps using a feedback network

Completing Partial Point Clouds with Outliers by Collaborative Completion and Segmentation

Skeleton-bridged Point Completion: From Global Inference to Local Adjustment

Shape Completion with Points in the Shadow

N-DPC: Dense 3D Point Cloud Completion Based on Improved Multi-Stage Network

SnowflakeNet: Point Cloud Completion by Snowflake Point Deconvolution with Skip-Transformer

Point Cloud Completion Via Skeleton-Detail Transformer

Learning Contours for Point Cloud Completion

A coarse-to-fine point completion network with details compensation and structure enhancement

Fast point completion network

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

Research on Multi-modal Point Cloud Completion Task