Search papers, labs, and topics across Lattice.
This study investigates the limitations of dense geometric prediction models in verifying physical causality by introducing "Scrambled Edges," a counterfactual that disrupts surface continuity and occlusion. The experiments reveal that these models exhibit a significant "Geometric Collapse," with deviations from clean predictions increasing by up to 3.2 times when faced with unsupported edge evidence. Notably, even with access to the corrupted regions, the models only recover 47% of the output, highlighting a critical gap in their ability to handle physically implausible cues.
Current dense predictors can deviate by over 3x from accurate predictions when faced with unsupported edge cues, revealing a fundamental flaw in their design.
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.
