Shape Completion Using Deep Boltzmann Machine

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Shape completion is a critical task in image processing.
• Generative models like Deep Boltzmann Machines (DBM) offer a powerful approach for shape completion by modeling shape distributions.
• Existing methods may require prior information about the incomplete object shape.

Purpose of the Study:

• To develop a novel shape completion method using Deep Boltzmann Machines (DBM).
• To improve the realism and efficiency of shape completion without requiring prior object information.
• To integrate convolutional shape features with DBM's generative capabilities.

Main Methods:

• Utilized the hidden activation of Deep Boltzmann Machines (DBM).
• Incorporated convolutional shape features with DBM hidden activations to create a regression model.
• Developed a method to compare regression model output with incomplete shape features for mask generation.
• Employed sampling from the DBM for shape completion.

Main Results:

• The proposed method successfully generates realistic shape completions.
• The technique does not require any prior information about the incomplete object shape.
• The integration of convolutional features and DBM activations proved effective for regression and sampling.

Conclusions:

• The developed method offers a robust and efficient solution for shape completion.
• This approach advances generative modeling for image processing tasks.
• The ability to complete shapes without prior information opens new possibilities in computer vision.