SLS4D: Sparse Latent Space for 4D Novel View Synthesis.

Area of Science:

• Computer Vision
• 3D Graphics
• Machine Learning

Background:

• Neural Radiance Fields (NeRF) excel at static scene novel view synthesis.
• Existing dynamic NeRF methods struggle with global dynamics and large model sizes.
• 4D scenes possess inherent spatial sparsity in deformation and temporal density.

Purpose of the Study:

• To develop a parameter-efficient method for dynamic Neural Radiance Fields (NeRF).
• To capture global dynamics in 4D scenes effectively.
• To improve 4D novel view synthesis performance.

Main Methods:

• Representing 4D scenes using a learnable sparse latent space (SLS4D).
• Utilizing dense time slot features to model temporal dynamics.
• Employing linear MLPs for deformation field prediction.
• Learning spatial features via an attention-based sparse latent space.

Main Results:

• SLS4D achieves state-of-the-art 4D novel view synthesis.
• The proposed method significantly reduces model parameters (approx. 6% of recent work).
• Demonstrates effective capture of global scene dynamics.

Conclusions:

• SLS4D offers a highly efficient and effective approach to dynamic NeRF.
• The sparse latent space representation addresses limitations of existing dynamic NeRF models.
• SLS4D advances the field of 4D scene representation and novel view synthesis.