Geometric robustness validation for end-to-end perception pipelines

Establish comprehensive geometric-robustness validation frameworks for multi-stage autonomous perception pipelines that account for continuous keypoint regression and downstream non-linear geometric solvers, rather than focusing exclusively on single-stage image classification.

Background

The paper considers safety-critical Vision-Based Landing systems composed of a neural network that detects two-dimensional runway keypoints and a Perspective-n-Point solver that estimates the aircraft’s three-dimensional pose. Physically plausible geometric perturbations, such as camera rotations, can alter early keypoint predictions and propagate non-linearly through the downstream solver, potentially producing large pose-estimation errors.

Existing robustness-validation frameworks are described as being primarily designed for single-stage image-classification models. The unresolved challenge is therefore to develop validation methods capable of assessing robustness across the complete perception chain, including continuous keypoint regression and non-linear geometric pose-estimation components.

References

Moreover, evaluating the geometric robustness of such multi-stage perception chains remains an open challenge; existing validation frameworks predominantly focus on single-stage image classification and fail to account for downstream tasks, such as non-linear geometric solvers.

— Robust Validation to Geometric Perturbations for Autonomous Pose Estimation  (2608.21066 - Theau et al., 21 Aug 2026) in Section 1, Introduction