Geometric Collapse: When Vision Models Fail to Verify Physical Causality

2026-07-07

AI Summary

Current vision models lack mechanisms to validate physical causal plausibility during inference, rendering them unable to detect implausible edges that violate surface continuity, illumination consistency, or occlusion ordering. This work proposes Scrambled Edges, a method that introduces physically implausible edge cues through controlled counterfactual perturbations while preserving energy and structural alignment, thereby disentangling the effects of high-frequency energy and edge sparsity for the first time. Evaluations on NYU Depth v2 and KITTI across diverse depth estimation architectures—including CNNs, Vision Transformers, self-supervised models, and diffusion or flow-matching frameworks—reveal that such perturbations induce prediction errors up to 3.2 times greater than those caused by energy-matched noise. Even when corrupted regions are known, output-level repair recovers only 47% of accuracy, with significant error propagation into unperturbed areas, demonstrating that existing models fail to effectively suppress physically invalid edge signals.

Abstract

Recent progress in large-scale self-supervised learning has improved dense geometric prediction, but it remains unclear whether such scaling yields inference-time physical plausibility checks. We propose Scrambled Edges, a controlled counterfactual that injects salient edge-like cues while violating surface continuity, illumination coherence, and occlusion ordering. With energy-matched and structure-matched controls, we isolate the effect of unsupported edge evidence from high-frequency energy and edge sparsity. Across CNN/ViT/SSL depth predictors on NYU Depth v2 and KITTI, Scrambled Edges induce up to 3.2x larger deviation from clean predictions than energy-matched noise; additional diffusion and flow-matching depth estimators show attenuated but still significant collapse. The resulting Geometric Collapse propagates globally: even with oracle knowledge of the corrupted region, output-level repair recovers only 47%, with substantial error outside the mask. These findings provide controlled behavioral evidence that current dense predictors lack reliable mechanisms to quarantine physically unsupported edge cues, motivating explicit plausibility scoring and selective cue integration.

Problem

• Geometric Collapse
• Physical Causality
• Dense Prediction
• Edge Cues
• Physical Plausibility

Innovation

• Geometric Collapse
• Scrambled Edges
• physical plausibility
• counterfactual perturbation
• dense depth prediction