Anisotropic Convolutional Neural Networks for RGB-D Based Semantic Scene Completion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Scene Understanding

Background:

• Semantic Scene Completion (SSC) infers scene occupancy and labels from partial visual data.
• Existing SSC methods struggle to effectively model complex visual and geometric scene variations.
• Voxel-wise labeling in SSC requires robust methods for handling incomplete scene information.

Purpose of the Study:

• To propose a novel network, Anisotropic Network (AIC-Net), for efficient and accurate semantic scene completion.
• To develop new convolutional modules capable of modeling anisotropic receptive fields for enhanced 3D context.
• To introduce an end-to-end trainable framework for SSC, bypassing costly pre-processing steps.

Main Methods:

• Introduced Anisotropic Network (AIC-Net) with novel anisotropic convolutional modules.
• Developed Kernel-Selection Anisotropic (KSA) and Kernel-Modulation Anisotropic (KMA) convolutions for adaptive receptive fields.
• Proposed an end-to-end trainable framework that avoids TSDF pre-processing.

Main Results:

• AIC-Net demonstrates effective modeling of visual and geometrical variations for scene completion.
• The anisotropic modules enhance 3D context modeling capability and computational efficiency.
• Experiments on standard SSC benchmarks validate the proposed method's advantages over existing approaches.

Conclusions:

• The proposed AIC-Net offers a significant advancement in semantic scene completion.
• Anisotropic convolutional modules provide a flexible and efficient way to model 3D scene context.
• The end-to-end framework simplifies the SSC pipeline and improves performance.