Hierarchical object parsing from structured noisy point clouds.

Area of Science:

• Computer Vision
• Machine Learning
• Computational Geometry

Background:

• Object parsing and segmentation from point clouds are difficult due to noisy data and thin structures.
• Existing models like Active Shape and Appearance Models (AAMs) lack flexibility for this task.
• Recursive compositional models simplify for computational guarantees, potentially losing accuracy.

Purpose of the Study:

• To investigate a hierarchical Bayesian model for shape and appearance in a generative setting.
• To develop a flexible model capable of accurately following object boundaries in noisy point cloud data.
• To introduce an efficient inference algorithm for improved object parsing.

Main Methods:

• A hierarchical Bayesian model is proposed, incorporating a generative approach.
• An object parsing layer deforms a hidden Principal Component Analysis (PCA) shape model with a Gaussian prior.
• A novel inference algorithm utilizes data-driven proposals to initialize local searches for hidden variables.

Main Results:

• The proposed model achieves state-of-the-art object parsing errors on two standard datasets.
• The approach successfully parses objects from structured point clouds, such as edge detection images.
• No intensity information was required, demonstrating the model's reliance on shape and structure.

Conclusions:

• The hierarchical Bayesian model offers a flexible and accurate solution for object parsing from challenging point cloud data.
• The novel inference algorithm enhances the efficiency and effectiveness of the parsing process.
• This method advances the state-of-the-art in point cloud object segmentation, particularly for boundary-focused data.