Research map / Adversarial robustness and security / Attacks on text and graphs
Targeted Adversarial Perturbations for Monocular Depth Prediction
Alex Wong, Safa Cicek, Stefano Soatto
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 |
|---|---|---|---|
| use proposed loss instead of DAG for scene scaling | targeted perturbation · KITTI scaling +-5%/+-10% | absolute relative error | Improved |
| increase perturbation norm xi from 2e-3 to 2e-2 (ablation) | targeted perturbation · KITTI scaling +-10% | absolute relative error: 3% -> 0.5% | Improved |
| horizontally flip instead of scaling by +10% | targeted perturbation · KITTI, xi=2e-2 | L1 norm of perturbation: 0.0161 -> 0.0326 | Hurt |
| vertical flip instead of horizontal flip | targeted perturbation · KITTI | attack success | Hurt |
| apply Gaussian blur defense at xi=2e-2 (ablation) | Monodepth2 · KITTI scaling +-10% | absolute relative error: 3% | Mixed |
| adversarially train Monodepth2 (ablation) | Monodepth2 · KITTI scaling +-10%, xi=2e-3 | absolute relative error: 7% | Mixed |
| adversarial training versus Gaussian blur at large xi (ablation) | defense · KITTI scaling | absolute relative error | Improved |
| adversarial training versus Gaussian blur at small xi (ablation) | defense · KITTI scaling | absolute relative error | No clear effect |
| constrain perturbation within instance mask (ablation) | targeted perturbation · Instance removal | removal success | Hurt |
| constrain perturbation outside instance mask (ablation) | targeted perturbation · Instance removal | removal success | Mixed |
| target traffic sign and human instead of vehicle and nature | category scaling · alpha=+-10% | absolute relative error: 1% -> 6% | Hurt |
| increase scale beyond 15% with xi=2e-3 (ablation) | targeted perturbation · larger scaling | absolute relative error: approximately 4% for every 5% increase | Hurt |
Closest papers by meaning
- Stereoscopic Universal Perturbations Across Different Architectures and Datasets (CVPR 2022 · similarity 0.91)
- Stereopagnosia: Fooling Stereo Networks with Adversarial Perturbations (AAAI 2021 · similarity 0.90)
- Semi-Supervised Adversarial Monocular Depth Estimation (TPAMI 2019 · similarity 0.89)
- Defending Against Universal Perturbations With Shared Adversarial Training (ICCV 2019 · similarity 0.88)
- On the Robustness of Language Guidance for Low-Level Vision Tasks: Findings from Depth Estimation (CVPR 2024 · similarity 0.86)
- DepthCues: Evaluating Monocular Depth Perception in Large Vision Models (CVPR 2025 · similarity 0.85)
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.
