IntrinsicNGP: Intrinsic Coordinate Based Hash Encoding for Human NeRF.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Neural radiance fields (NeRFs) are effective for novel view synthesis but typically require extensive training times.
• Existing methods struggle with dynamic scenes and practical applications due to long training durations.

Purpose of the Study:

• To develop a fast and efficient method for high-fidelity novel view synthesis of human performers.
• To enable practical applications of neural radiance fields by significantly reducing training time.

Main Methods:

• Introduced IntrinsicNGP, a novel approach utilizing continuous and optimizable intrinsic coordinates within the hash encoding module.
• Employed proxy geometry shapes to aggregate interframe information for dynamic objects.
• Incorporated an optimizable offset field, guided by intrinsic coordinates, for refining results.

Main Results:

• Achieved high-fidelity novel view synthesis results in minutes, a significant improvement over existing methods.
• Demonstrated effectiveness and efficiency across multiple datasets.
• Showcased the ability to edit the shape of reconstructed objects.

Conclusions:

• IntrinsicNGP offers a practical and efficient solution for neural radiance field training.
• The proposed intrinsic coordinate system enhances dynamic scene reconstruction and allows for shape manipulation.