Cascaded Refinement Network for Point Cloud Completion With Self-Supervision.

Area of Science:

• Computer Vision
• 3D Geometry Processing
• Machine Learning

Background:

• Point clouds often suffer from sparsity and incompleteness, hindering real-world applications.
• Existing shape completion techniques frequently produce coarse shapes lacking fine-grained details.
• Acquiring large amounts of ground truth data for training is challenging in practical scenarios.

Purpose of the Study:

• To develop an advanced two-branch network for high-fidelity point cloud shape completion.
• To improve the preservation of object details during dense point reconstruction.
• To reduce the reliance on extensive ground truth data through novel training strategies.

Main Methods:

• A two-branch network architecture: one for cascaded shape completion and another as an auto-encoder.
• Shared feature extractor to learn accurate global features for both branches.
• Proposed two training strategies to enable learning without ground truth data, enhancing self-supervised and semi-supervised learning.

Main Results:

• The cascaded completion branch utilizes partial input and coarse output for detailed reconstruction.
• The proposed training strategies improve reconstruction quality in fully supervised settings.
• Superior performance demonstrated across self-supervised, semi-supervised, and fully supervised learning scenarios.

Conclusions:

• The developed method achieves more realistic outputs compared to state-of-the-art approaches for point cloud completion.
• The approach effectively handles sparse and incomplete point cloud data, preserving fine-grained details.
• The novel training strategies mitigate the need for large ground truth datasets, increasing practical applicability.