Diffusion-Driven Self-Supervised Learning for Shape Reconstruction and Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Fully-supervised category-level pose estimation requires costly manual annotations.
• Existing self-supervised methods often rely on synthetic data or CAD models and are limited to single objects.
• Multi-object pose estimation and shape reconstruction remain challenging tasks.

Purpose of the Study:

• To develop a diffusion-driven self-supervised network for multi-object shape reconstruction and category-level pose estimation.
• To overcome limitations of existing methods by leveraging shape priors without synthetic data or CAD models.
• To address challenges in intra-class shape variations and improve performance on complex scenes.

Main Methods:

• Introduced a Prior-Aware Pyramid 3D Point Transformer for SE(3)-equivariant pose features and 3D scale-invariant shape information.
• Utilized point convolutional layers with radial-kernels for pose-aware learning and graph convolution for shape representation.
• Developed a Pretrain-to-Refine Self-SuperVISED Training Paradigm incorporating a diffusion mechanism.

Main Results:

• The proposed network achieved state-of-the-art performance in self-supervised category-level pose estimation.
• Outperformed existing self-supervised methods on multiple public datasets and a custom dataset.
• Demonstrated effectiveness in multi-object scenarios and shape reconstruction tasks, surpassing some fully-supervised methods.

Conclusions:

• The diffusion-driven self-supervised network offers a powerful solution for category-level pose estimation and shape reconstruction.
• The method effectively leverages shape priors and handles intra-class variations, reducing reliance on manual annotations.
• This approach advances self-supervised learning in computer vision for complex 3D scene understanding.