Recurrent Diffusion for 3D Point Cloud Generation From a Single Image.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Generative models, particularly diffusion models, have advanced single-image 3D shape reconstruction.
• Existing methods suffer from cumulative errors and detail loss due to single forward passes.
• Limited input information from single images degrades performance in current 3D reconstruction techniques.

Purpose of the Study:

• To develop a recurrent diffusion framework for improved single-image 3D shape reconstruction.
• To address cumulative errors and enhance geometric consistency in generated 3D shapes.
• To boost the performance of 3D shape generation using only a single input image.

Main Methods:

• A recurrent diffusion framework recursively refines noise prediction with target guidance.
• A self-rectified approach suppresses cumulative errors and improves detail modeling.
• A multi-view training scheme with view-robust conditional generation enhances geometric perception for single-image inference.

Main Results:

• The proposed method significantly suppresses cumulative errors in 3D shape reconstruction.
• Detail modeling and geometric consistency are markedly improved compared to prior works.
• The framework achieves state-of-the-art performance on public 3D shape datasets, both qualitatively and quantitatively.

Conclusions:

• The recurrent diffusion framework offers a robust solution for single-image 3D shape reconstruction.
• The multi-view training scheme effectively enhances network perception for limited input scenarios.
• This work advances the quality and reliability of generative models for 3D shape generation from single images.