Semantic Scene Completion Using Local Deep Implicit Functions on LiDAR Data.

Area of Science:

• Computer Vision
• 3D Scene Understanding
• Machine Learning

Background:

• Semantic scene completion aims to reconstruct 3D geometry and semantics from sparse, occluded real-world data.
• Existing methods often rely on voxelization, limiting scene detail or extent.

Purpose of the Study:

• To develop a novel learning-based method for semantic scene completion.
• To overcome limitations of voxel-based representations by proposing a continuous scene representation.

Main Methods:

• A scene segmentation network utilizing local Deep Implicit Functions is proposed.
• Raw point clouds are encoded into a multi-resolution latent space locally.
• A global scene completion function is assembled from localized function patches.

Main Results:

• The method produces a continuous scene representation suitable for extensive outdoor scenes.
• This representation effectively encodes geometric and semantic properties without spatial discretization.
• Performance surpasses state-of-the-art on the Semantic KITTI Scene Completion Benchmark in geometric completion IoU.

Conclusions:

• The proposed method generates a powerful, continuous scene representation for 3D scene completion.
• It effectively decodes into a dense 3D scene description, advancing the state-of-the-art.