Structure from Articulated Motion: Accurate and Stable Monocular 3D Reconstruction without Training Data.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Motion Capture

Background:

• Recovering 3D articulated structures from 2D images is a complex computer vision task.
• Existing learning-based methods excel on benchmarks but lack generalizability.
• Model-based methods are general but less accurate.

Purpose of the Study:

• Introduce a novel model-based method, Structure from Articulated Motion (SfAM).
• Enable recovery of diverse object and motion types without reliance on large training datasets.
• Achieve state-of-the-art accuracy comparable to learning-based methods.

Main Methods:

• Developed SfAM, a general-purpose non-rigid structure from motion (NRSfM) technique.
• Integrated a soft spatio-temporal constraint on bone lengths.
• Employed alternating optimization for joint positions and bone proportions recovery.

Main Results:

• SfAM performs on par with state-of-the-art learning-based approaches on public benchmarks.
• SfAM outperforms previous non-rigid structure from motion (NRSfM) methods.
• Demonstrated robustness to noisy 2D data and generalization to arbitrary objects.

Conclusions:

• SfAM offers a robust, data-efficient approach to monocular 3D recovery of articulated structures.
• The method generalizes across object types and motions without task-specific training.
• SfAM provides a new perspective for applications like human motion capture.