Download PDF.pdf

This is a novel contribution to the field of computer vision, addressing challenges posed by increasingly common scenarios involving partially overlapping point data, like those from LiDAR systems or SLAM applications.

Key Issues:

Existing point cloud processing methods lack specific consideration for partially overlapping point clouds. While 3D point cloud segmentation techniques exist (e.g., Qi et al., 2017b; Tchapmi et al., 2017; Wang et al., 2018), they focus on single point cloud segmentation, not overlapping region identification between multiple clouds.

Proposed Solution:

POCS-Net comprises three main components:

1. Local Feature Extractor: Extracts rotation-invariant local features from each point in the point cloud.

2. Co-Attention Module: Utilizes a co-attention mechanism to integrate local features from both overlapping point clouds, prioritizing information related to the overlap and distinguishing it from non-overlapping regions.

3. Segmentation Module: Employs a fully connected network to predict the probability that each point belongs to the overlapping region.

Training & Evaluation:

POCS-Net is trained using a novel dataset constructed from ModelNet40, ShapeNet, 3DMatch, and ScanNet, encompassing partially overlapping point clouds obtained automatically or through relative pose information.
The network is trained to maximize similarity between local features of the overlap region while encouraging larger overlap areas.
* Two regularization techniques are implemented: a sphere distribution regulation term and sharing the encoder with a self-supervised learning task to prevent feature degeneration.

Downstream Application:

The paper demonstrates the effectiveness of POCS-Net in improving 3D point cloud registration accuracy and robustness by segmenting overlapping regions, making downstream tasks easier to solve.