${\rm{H}}_{2}{\rm{O}}$H2O-NeRF: Radiance Fields Reconstruction for Two-Hand-Held Objects.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Neural Rendering

Background:

• Reconstructing two-hand-held objects presents challenges due to complex occlusions and view-dependent effects.
• Traditional single-hand-held object reconstruction methods are limited in flexibility and completeness.
• Neural rendering offers promising avenues for advanced 3D object reconstruction.

Purpose of the Study:

• To develop a novel neural representation-based framework for reconstructing the appearance and geometry of two-hand-held objects.
• To address the complex occlusions inherent in two-hand manipulation for improved 3D reconstruction.
• To enhance the completeness and view-consistency of reconstructed objects.

Main Methods:

• Proposed HO-NeRF framework utilizing an object-centric semantic module and a view-dependent visible module.
• Integration of geometric signed distance function cues for 3D region prediction.
• Development of a 2D visible mask to guide adaptive ray sampling during optimization.
• Utilized a newly collected HO dataset for method validation.

Main Results:

• HO-NeRF demonstrates superior performance in reconstruction completeness compared to state-of-the-art methods.
• The framework achieves enhanced view-consistency synthesis in rendered outputs.
• Experimental validation on a novel HO dataset confirms the method's effectiveness.

Conclusions:

• HO-NeRF provides an effective solution for the challenging problem of two-hand-held object reconstruction.
• The proposed modules successfully mitigate complex occlusions, leading to more complete 3D models.
• This work advances the field of neural rendering for practical 3D object reconstruction applications.