Research map / Adversarial robustness and security / Attacks on text and graphs
Stereoscopic Universal Perturbations Across Different Architectures and Datasets
Zachary Berger, Parth Agrawal, Tian Yu Liu, Stefano Soatto, Alex Wong
Check 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 |
|---|---|---|---|
| Add an image-agnostic SUP optimized for the attacked model | AANet, DeepPruner, and PSMNet · KITTI 2015 | D1-error: 1.47% -> 48.43%; 1.28% -> 52.74%; 4.25% -> 87.72% | Hurt |
| Transfer AANet-optimized SUPs across architectures | DeepPruner and PSMNet · KITTI 2015 | D1-error: 1.28% to 52.66%; 4.25% to 61.66% | Hurt |
| Apply KITTI-optimized SUPs to models trained on Scene Flow | AANet, DeepPruner, and PSMNet · FlyingThings3D | EPE: 1.30px to 9.47px; 1.25px to 14.77px; 1.27px to 18.88px | Hurt |
| Use a 64 x 64 tiled SUP instead of a full-image SUP (ablation) | AANet, DeepPruner, and PSMNet · FlyingThings3D | reported error: 36.09% -> 46.14%; 23.28% -> 34.87%; 25.35% -> 31.93% | Hurt |
| Use a 64 x 64 tile instead of a 128 x 128 tile (ablation) | AANet, DeepPruner, and PSMNet · FlyingThings3D | reported error: 40.84% -> 46.14%; 27.89% -> 34.87%; 29.33% -> 31.93% | Hurt |
| Increase the tile size from 64 x 64 to 128 x 128 (ablation) | AANet · KITTI 2015 | D1-error | Improved |
| Add deformable convolutions (ablation) | DeepPruner and PSMNet · KITTI 2015 | D1-error: 52.74% -> 39.47%; 87.72% -> 52.10% | Improved |
| Replace deformable convolutions with regular convolutions (ablation) | AANet · KITTI 2015 | D1-error: 48.43% -> 54.32% | Hurt |
| Increase deformable convolutions from 6 to 25, then add explicit matching (ablation) | PSMNet · KITTI 2015 | D1-error: 87.72% to 66.10%, 52.10%, and finally to 39.47% | Improved |
| Fine-tune with adversarial data augmentation (ablation) | AANet, DeepPruner, and PSMNet · KITTI 2015 | D1-error: 48.43% to 3.62%; 52.74% to 2.83%; 87.72% to 2.96% | Improved |
| Fine-tune with adversarial data augmentation against a newly optimized adversary (ablation) | AANet, DeepPruner, and PSMNet · KITTI 2015 | D1-error: 25.54%; 31.54%; 35.75% | Mixed |
| Combine 25 deformable convolutions and explicit matching (ablation) | PSMNet · KITTI 2015 | D1-error: 87.72% -> 39.47% | Improved |
Closest papers by meaning
- Stereopagnosia: Fooling Stereo Networks with Adversarial Perturbations (AAAI 2021 · similarity 0.95)
- Targeted Adversarial Perturbations for Monocular Depth Prediction (NeurIPS 2020 · similarity 0.91)
- Revisiting Non-Parametric Matching Cost Volumes for Robust and Generalizable Stereo Matching (NeurIPS 2022 · similarity 0.90)
- Cheating Stereo Matching in Full-Scale: Physical Adversarial Attack Against Binocular Depth Estimation in Autonomous Driving (AAAI 2026 · similarity 0.88)
- DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches (NeurIPS 2025 · similarity 0.88)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.87)
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.
