Scribble-Based 3D Shape Segmentation via Weakly-Supervised Learning.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Shape segmentation is crucial in 3D shape analysis.
• Prior knowledge enhances segmentation accuracy but requires extensive manual labeling.
• Fully labeled datasets are labor-intensive to create.

Purpose of the Study:

• To develop a novel weakly-supervised algorithm for 3D shape segmentation.
• To reduce the manual workload associated with creating labeled training data.
• To improve segmentation accuracy using partial, scribble-based labeling.

Main Methods:

• A deep learning approach that jointly propagates information from scribbles to unlabeled faces.
• Simultaneous learning of deep neural network parameters and label propagation.
• Utilizes a scribble-based, partially supervised labeling process.

Main Results:

• The proposed method achieves superior segmentation performance compared to unsupervised approaches.
• It demonstrates comparable segmentation performance to state-of-the-art fully supervised methods.
• Experimental results validate the effectiveness of the weakly-supervised approach.

Conclusions:

• Weakly-supervised learning significantly reduces the burden of 3D shape segmentation data labeling.
• The proposed algorithm offers an efficient and effective alternative to fully supervised methods.
• This approach enables accurate 3D shape segmentation with minimal manual input.