Iterate or Widen? When Test-Time Refinement Helps LiDAR Scene Completion: A Controlled Study of Evidence Geometry, Training Coverage, and Compute

2026-08-06

AI Summary

This study addresses the unresolved question in LiDAR semantic scene completion of whether test-time iterative refinement or increasing model width yields better performance. Through rigorously controlled experiments with matched computational budgets, the authors compare single-pass prediction, width-scaled models, and a weight-sharing multi-grid iterative optimizer built upon a frozen backbone, evaluating robustness across controlled point cloud degradation modes—including coherent occlusion, independent sparsification, range-based attenuation, and clutter. Results demonstrate that iterative optimization provides substantial gains only under coherent missing regions (e.g., +0.911 mIoU on SemanticKITTI sequence 08, 95% CI [0.804, 1.040]), offers marginal improvement under sparse missingness (+0.300 points), and fails to handle clutter, at a cost of 10.74 ms and 0.75 GiB per frame. This work is the first to reveal the geometric dependency of iterative strategies under consistent training–testing conditions, offering empirical guidance for test-time compute allocation.

Abstract

Should a completion model spend extra test-time compute by iterating, or spend a similar parameter budget on a wider one-shot predictor? The answer is easily confounded by denoising curricula, corruption augmentation, capacity, and unpaired evaluation. We study this question in LiDAR semantic scene completion by comparing a one-shot predictor, a parameter-matched wider predictor, and a weight-tied multigrid refiner initialized from the same frozen predictor. The protocol separates coherent region removal, independent thinning, range-dependent attenuation, and additive clutter while preserving exact scene-condition pairing. Across five training seeds and 815 SemanticKITTI sequence-08 frames, the full iterative system improves mIoU over the wide control by 0.911 points under contiguous angular removal, with a 95% moving-block bootstrap interval of [0.804, 1.040] that clears a predeclared 0.5-point practical margin. Under independent 75% thinning, iteration adds only 0.300 points [0.166, 0.436], whereas observation-family augmentation adds 5.975 points [5.662, 6.140]. Neither intervention repairs additive clutter. The iterative system also costs 10.74 ms and 0.75 GiB per frame, versus 6.25 ms and 0.23 GiB for the wide control. These results establish a geometry-conditioned empirical boundary rather than a universal advantage: coherent gaps can justify fixed-depth refinement, broadly thinned evidence is addressed more effectively by training coverage, and spurious evidence requires a different robustness mechanism.

Problem

• LiDAR scene completion
• test-time refinement
• iterative vs one-shot prediction
• evidence geometry
• training coverage

Innovation

• test-time refinement
• LiDAR scene completion
• controlled ablation study
• evidence geometry
• parameter-matched comparison