ANISE: Assembly-Based Neural Implicit Surface Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• 3D shape reconstruction from limited data is a challenging problem.
• Existing methods often struggle with part-aware representations and coarse-to-fine reconstruction strategies.

Purpose of the Study:

• To introduce ANISE, a novel method for 3D shape reconstruction from partial observations.
• To leverage a part-aware neural implicit shape representation for improved reconstruction accuracy.
• To enable reconstruction through both direct decoding and part database assembly.

Main Methods:

• ANISE formulates 3D shapes as assemblies of neural implicit functions, each representing a part instance.
• The reconstruction process follows a coarse-to-fine approach, predicting structural arrangements before detailed geometry.
• Part latent codes are generated, enabling reconstruction via decoding to implicit functions or retrieval from a part database.

Main Results:

• ANISE achieves state-of-the-art performance in part-aware reconstruction from both images and sparse point clouds when decoding part representations.
• Reconstructing shapes by assembling retrieved parts significantly outperforms traditional methods, even with limited database sizes.
• The method's effectiveness is validated on standard sparse point cloud and single-view reconstruction benchmarks.

Conclusions:

• ANISE offers a robust and effective approach to 3D shape reconstruction from partial observations.
• The part-aware neural implicit representation and coarse-to-fine strategy are key to its superior performance.
• The method demonstrates flexibility through its dual reconstruction pathways, enhancing its applicability.