Parameterization of Volume Sampling for Active Learning of Radiance Field.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Radiance field techniques excel at photorealistic scene reconstruction.
• A key limitation is the need for densely captured training images.
• Active learning can optimize image acquisition for efficiency.

Purpose of the Study:

• To develop a computationally tractable method for estimating information gain in active learning for radiance fields.
• To improve the efficiency and effectiveness of view selection for radiance field training.

Main Methods:

• Leveraging volume rendering's ray sampling process to approximate expected information gain.
• Parameterizing volume sampling with perturbed ray directions.
• Learning the predictive distribution to infer optimal perturbation patterns for view selection.

Main Results:

• Achieved state-of-the-art performance in image quality for active learning of radiance fields.
• Outperformed previous methods across diverse datasets (forward-facing and object-centric).
• Demonstrated effective exploration of volume sampling space for enhanced diversity and informative view selection.

Conclusions:

• The proposed method offers a computationally efficient and effective solution for active learning in radiance fields.
• This approach enhances the diversity and informativeness of selected views, leading to superior scene reconstruction.
• The technique shows broad applicability across different types of scenes.