AMVP: Adaptive Multi-Volume Primitives for Auto-Driving Novel View Synthesis

Synthesizing high-quality novel views is critical to extending training data for auto-driving scenes. However, existing novel view synthesis techniques rely on a single-volume radiance field with uniform spatial resolution, constraining their model capacity and resulting in artifacts in synthesized auto-driving views. This letter introduces AMVP, a novel neural representation that models auto-driving scenes using multiple local primitives with adaptive spatial resolution. AMVP addresses the lack of representation capability of detail-rich regions by adaptively subdividing the scene into multiple local volumes. Each local volume is assigned a tailored resolution based on its geometric complexity, as determined by a density prior. Subsequently, multi-volume primitives are introduced to enable sharing a global feature table among local volumes, addressing the GPU memory inefficiency caused by the duplicated allocation. In addition, the letter proposes resolution-aware confidence, a mechanism that suppresses artifacts arising from frequency ambiguity. This mechanism adaptively reduces high-frequency components based on the spatial resolution of each local volume and the distance of the sampling point from the optical center. Experimental results on benchmark auto-driving datasets demonstrate that the proposed AMVP achieves superior rendering quality while using a similar number of parameters compared to existing methods.