Exploring Explicit Representations in 4D: A Comparative Analysis with HexPlane

TMLR Paper2286 Authors

Abstract:

Modeling and re-rendering novel views of dynamic 3D scenes is a challenging problem in
3D vision. Employing implicit representations for the task, extending static NeRFs to
4D incurs high computational costs due to the numerous MLP evaluations, highlighting
the need for efficient representations of dynamic 3D scenes. Non-Nerf Methods such as
Niemeyer et al. (2019), Jiang et al. (2022), and Jiang et al. (2021) have primarily been applied
to idealized, single-subject scenes and have not yet been adapted for real-world camera
images. Cao and Johnson (2023) proposes using HexPlane, an explicit scene representation
method that factors a 4D volume into six feature planes. This paper attempts to verify
their claims and compare them with similar methods like Gaussian Splatting by Wu et al.
(2023) and K-planes by Fridovich-Keil et al. (2023). We conduct a thorough examination of
the architectural choices and design elements inherent in HexPlane and further incorporate
additional regularization to achieve a performance improvement.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Yanwei_Fu2

Submission Number: 2286