Monocular 3D Body Shape Reconstruction under Clothing.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Human Body Modeling

Background:

• Estimating 3D object shape from monocular images is a complex computer vision task.
• Accurate 3D human body shape estimation from single images is particularly challenging due to clothing obscuring the underlying form.

Purpose of the Study:

• To develop a robust method for estimating 3D human body shape from single images, even when the subject is wearing loose clothing.
• To leverage a parametric 3D body model (SMPL) and deep learning to infer body shape parameters.

Main Methods:

• Utilized the SMPL (Skinned Multi-Person Linear) model to represent body pose and shape.
• Trained a deep convolutional neural network to regress SMPL shape parameters from single images.
• Developed a training dataset by associating minimally clothed body shape parameters with images of the same person in looser clothing to enhance robustness.

Main Results:

• The proposed deep learning approach demonstrated superior accuracy in estimating 3D body shape parameters compared to existing state-of-the-art methods.
• The method proved effective even in challenging scenarios involving loose clothing that significantly hides the body's true shape.

Conclusions:

• The developed method offers a significant advancement in 3D human body shape estimation from monocular images, particularly under clothing.
• The strategy of training with varied clothing conditions enhances the model's robustness and accuracy for real-world applications.