Latent-Class Hough Forests for 6 DoF Object Pose Estimation

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Object detection and pose estimation are challenging in cluttered and occluded environments.
• Existing methods struggle with heavy clutter and partial occlusions, limiting real-world applicability.

Purpose of the Study:

• To introduce Latent-Class Hough Forests (LCHF) for robust object detection and 6 Degrees of Freedom (6 DoF) pose estimation.
• To enhance detection rates and provide occlusion-aware segmentation in complex scenarios.

Main Methods:

• Adapted state-of-the-art template matching features into scale-invariant patch descriptors.
• Integrated descriptors into a regression forest with a novel template-based split function.
• Treated class distributions as latent variables, inferring them iteratively during testing.

Main Results:

• Achieved accurate estimation of background clutter and foreground occlusions, improving detection rates.
• Generated accurate occlusion-aware segmentation masks, even for multi-instance scenarios.
• Outperformed state-of-the-art methods on public and newly collected challenging datasets.

Conclusions:

• Latent-Class Hough Forests offer a significant advancement in object detection and pose estimation under occlusion.
• The method demonstrates robustness in complex, real-world scenarios with heavy clutter.
• LCHF provides valuable by-products like occlusion-aware segmentation masks.