Fast Non-Rigid Radiance Fields from Monocularized Data.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Reconstructing dynamic scenes from large-scale multi-view data is computationally intensive.
• Current methods using monocularized data struggle with training speed and limited angular range for novel view synthesis.
• Existing techniques often fail to address inward-facing dynamic scenes effectively.

Purpose of the Study:

• To develop a novel method for full 360° inward-facing novel view synthesis of non-rigidly deforming scenes.
• To address limitations in training speed and angular range of existing approaches for dynamic scene reconstruction.
• To enable efficient and accurate novel view synthesis from monocularized data.

Main Methods:

• Proposed an efficient deformation module decoupling spatial and temporal information for accelerated processing.
• Utilized a static module representing the canonical scene as a fast hash-encoded neural radiance field.
• Systematically analyzed performance on real-world inward-facing scenes using a newly recorded dataset.

Main Results:

• Achieved significantly faster convergence (under 7 minutes) compared to previous methods.
• Enabled real-time novel view synthesis at 1K resolution.
• Demonstrated higher visual accuracy for generated novel views on both synthetic and real-world data.

Conclusions:

• The proposed method offers a significant improvement in speed and accuracy for 360° novel view synthesis of dynamic scenes.
• The efficient deformation and static modules are key to accelerating training and inference.
• The method shows strong performance on challenging real-world inward-facing datasets, outperforming existing approaches.