Object-Occluded Human Shape and Pose Estimation With Probabilistic Latent Consistency.

Area of Science:

• Computer Vision
• 3D Human Pose and Shape Estimation

Background:

• Object occlusions are prevalent in human-object interaction scenarios, posing challenges for existing 3D human estimation methods.
• Current approaches often fail when human bodies are significantly occluded by objects.

Purpose of the Study:

• To develop a novel method for directly estimating 3D human shape and pose from single color images, even with substantial object occlusions.
• To address the limitations of existing methods in handling occluded human bodies.

Main Methods:

• Proposes a two-branch network architecture utilizing a partial UV map to represent occluded human bodies.
• Frames the full 3D human shape estimation as an image inpainting problem.
• Introduces a visible feature sub-network to extract human information from occluded images and employs latent distribution consistency for training.

Main Results:

• The proposed method effectively estimates 3D human shape and pose from object-occluded images.
• Achieves state-of-the-art performance compared to existing methods on this challenging task.
• Introduced a new dataset (3DOH50K) for training and evaluating occluded human shape and pose estimation.

Conclusions:

• The novel approach successfully reconstructs 3D human shape and pose despite significant object occlusions.
• The image inpainting formulation and specialized network architecture demonstrate effectiveness.
• The developed dataset and method offer significant advancements in handling occluded human data.