3D Human Pose, Shape and Texture From Low-Resolution Images and Videos.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Human Pose and Shape Estimation

Background:

• Estimating 3D human pose and shape from monocular images is crucial but challenging with low-resolution inputs.
• Current deep learning methods struggle with low-resolution data due to artifacts from super-resolution or the impracticality of training models for each resolution.

Purpose of the Study:

• To develop a single, robust algorithm capable of estimating 3D human pose and shape across diverse image resolutions.
• To address the limitations of existing methods in handling low-resolution monocular images for 3D human analysis.

Main Methods:

• Proposes RSC-Net, integrating a Resolution-aware network, Self-supervision loss, and Contrastive learning scheme.
• Self-supervision loss ensures output scale-consistency; contrastive learning ensures deep feature scale-consistency.
• Extends RSC-Net for low-resolution video analysis and textured 3D pedestrian reconstruction.

Main Results:

• RSC-Net learns 3D body pose and shape effectively across different resolutions using a single model.
• The proposed losses enhance robustness in weakly-supervised learning scenarios.
• Achieves superior performance compared to state-of-the-art methods on challenging low-resolution images.

Conclusions:

• RSC-Net offers a unified and effective solution for 3D human pose and shape estimation from low-resolution images and videos.
• The novel approach demonstrates significant improvements in accuracy and robustness for real-world applications.