Research map / Architectures and efficiency / RNNs and normalization
Geometric Collapse: When Vision Models Fail to Verify Physical Causality
Wentao Zhang, Jinhu Qi, Weiqiang Jin, Yifei Zhang, Chan-Tong Lam, Irwin King
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| violate occlusion causality (Direction cue) in edge perturbation (ablation) | MiDaS DPT · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 2.22x vs 1.00x for edge-shaped noise | Hurt |
| violate occlusion causality (Direction cue) (ablation) | DepthAnything V2 · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 3.18x | Hurt |
| apply structure-only edge-shaped noise (ablation) | DepthAnything V2 · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 1.71x | Mixed |
| place noise in the scrambled mask without edge structure (ablation) | all four depth models · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 0.58x, 0.60x, 0.43x, 0.53x | No clear effect |
| combine all prior violations (Scrambled Edges) versus Direction only (ablation) | DepthAnything V2 · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 3.18x -> 3.20x | Improved |
| increase perturbation intensity alpha from 0.2 to 0.8 (ablation) | MiDaS DPT · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 1.37x -> 2.34x | Improved |
| increase number of edge segments K from 5 to 20 (ablation) | MiDaS DPT · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 2.10x -> 2.39x | Improved |
| swap feed-forward estimator for diffusion (Marigold) | depth estimator · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 2.34x -> 1.55x | Improved |
| swap feed-forward estimator for flow matching (DepthFM) | depth estimator · NYU Depth v2 (N=1449) | collapse ratio vs high-pass noise: 2.34x -> 1.11x | Improved |
| oracle knowledge of perturbed pixels at the output instead of the input | MiDaS DPT · NYU Depth v2 (N=1449) | recovery (%): 99.96% -> 47.0% | Hurt |
| inpaint with oracle mask at input versus output | MiDaS DPT · NYU Depth v2 (N=1449) | recovery (%): 24.0% vs 21.0% | Mixed |
| add edge-consistency post-hoc defense | DepthAnything V2 · NYU Depth v2 (N=1449) | recovery (%): 0.2252 -> 0.1560 RMSE vs clean; 26.4% recovery | Improved |
Closest papers by meaning
- MEMO: Human-like Crisp Edge Detection Using Masked Edge Prediction (CVPR 2026 · similarity 0.86)
- GEOPHYS: The Geometry of Physical Plausibility (NeurIPS 2026 · similarity 0.86)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.85)
- The Impact of Geometric Complexity on Neural Collapse in Transfer Learning (NeurIPS 2024 · similarity 0.84)
- FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · similarity 0.84)
- Directional Confusions Reveal Divergent Inductive Biases Through Rate-Distortion Geometry in Human and Machine Vision (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
