SAND: Spatially Adaptive Network Depth for Fast Sampling of Neural Implicit Surfaces
Researchers have introduced SAND, a novel framework for neural implicit surfaces designed to optimize computational efficiency. SAND adaptively adjusts network depth based on spatial complexity and accuracy requirements, reducing wasted computations. This approach utilizes a volumetric depth map and a modified multi-layer perceptron to allow evaluations to terminate early in less complex regions, thereby speeding up inference while maintaining high-fidelity representations. AI
IMPACT Improves inference speed for neural implicit representations by adaptively adjusting network depth.