Resolution Where It Counts: Hash-based GPU-Accelerated 3D Reconstruction via Variance-Adaptive Voxel Grids

Lorenzo De Rebotti, Emanuele Giacomini, Giorgio Grisetti, Luca Di Giammarino

Abstract:

Efficient 3D surface reconstruction from range data in real-time scenarios remains computationally and memory-intensive. Traditional volumetric methods using fixed grids or octrees are memory-inefficient, computationally expensive, and lack full GPU support.
We present a variance-adaptive, multi-resolution voxel grid that adjusts resolution based on local SDF variance using a flat spatial hash table that enables constant-time access and full GPU parallelism. The system includes GPU-accelerated rendering via parallel quadtree-based Gaussian Splatting with adaptive splat density control.
Our CUDA/C++ implementation achieves up to 5× speedup and 4× memory reduction versus fixed-resolution methods with only 3.3 % less accuracy on average.

Submission Number: 9