Structural Outlier-Aware Post-Training Quantization for Monocular Depth Estimation

ViT-based monocular depth estimation (MDE) models achieve strong accuracy but remain costly to deploy, motivating post-training quantization (PTQ). Existing PTQ methods are not well aligned with dense depth prediction, and recent depth-specific methods provide limited insight into which operators fail and why. We analyze 4-bit PTQ failure across operators using output sensitivity to clipping and cross-input range stability, showing that activation outliers play three distinct roles: range-dominating, signal-bearing, and input-dependent. Based on this analysis, we propose ORA-Q, an operator-role-aware PTQ framework that assigns each operator its grouping granularity, calibrator, and static or dynamic scaling mode. ORA-Q keeps most operators static and applies dynamic scaling only to input-dependent ranges. Experiments on Depth Anything variants and depth benchmarks show that ORA-Q consistently outperforms prior PTQ methods, improving $\delta_1$ by 0.106 on average under 4-bit quantization.