Learning to Detect 3D Symmetry From Single-View RGB-D Images With Weak Supervision.

Area of Science:

• Computer Vision
• 3D Geometry
• Machine Learning

Background:

• 3D symmetry detection is crucial for computer vision and graphics.
• Existing methods often require complete 3D models or extensive symmetry annotations.
• Detecting symmetry from partial observations, like single RGB-D images, remains challenging.

Purpose of the Study:

• To develop a 3D symmetry detection method for single-view RGB-D images without symmetry supervision.
• To enable accurate symmetry detection from incomplete and partially observed 3D data.
• To overcome limitations of tedious manual annotation in existing approaches.

Main Methods:

• A weakly-supervised learning approach is employed, training a network to complete shapes based on predicted symmetries.
• A discriminative variational autoencoder learns a shape prior to assess 3D shape plausibility.
• A novel symmetry parametrization facilitates end-to-end learning for both reflectional and rotational symmetries.

Main Results:

• The proposed method accurately detects reflectional and rotational symmetries from single RGB-D images.
• It demonstrates robustness in scenarios with occlusion and scanning noise.
• Achieved state-of-the-art performance, outperforming supervised methods by 2%-11% F1-score on ShapeNet and ScanNet.

Conclusions:

• Coupling symmetry detection with shape completion enables learning a symmetry-aware shape prior.
• The approach provides accurate and robust 3D symmetry detection without supervision.
• This method advances the field by addressing the practical challenge of limited annotated data.