Lightweight 3-D Convolutional Occupancy Networks for Virtual Object Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Geometric Deep Learning

Background:

• Edge devices require adaptable technologies for demanding tasks like 3-D object reconstruction.
• Existing methods often prioritize mesh quality over computational speed, limiting their use on embedded systems.

Purpose of the Study:

• To develop a lightweight implicit representation for efficient 3-D object reconstruction.
• To achieve a balance between mesh generation time and mesh quality for embedded applications.

Main Methods:

• Leveraging convolutional occupancy networks for lightweight implicit representation.
• Implementing and testing the approach on the ShapeNet dataset.

Main Results:

• Demonstrated the effectiveness of the proposed approach on diverse hardware, including GPUs, CPUs, and embedded devices.
• Achieved a favorable balance between mesh generation speed and quality.

Conclusions:

• The proposed convolutional occupancy network approach offers an efficient solution for 3-D object reconstruction on resource-constrained edge devices.
• This method advances the feasibility of complex computer graphics tasks in augmented and virtual reality on non-specialized hardware.