Research map / 3D vision / Human pose estimation
Strike (With) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Search familiar objects across unrestricted 6D poses | Inception-v3 on the 30-object dataset · 6D pose space | correct classification rate: 3.09% | Hurt |
| Use finite-difference gradient descent instead of ZRS | Targeted pose optimization · Targeted adversarial poses | hit rate: 78 -> 92 | Helped |
| Use differentiable-renderer gradients instead of ZRS | Targeted pose optimization · Targeted adversarial poses | hit rate: 78 -> 32 | Hurt |
| Train with adversarial poses from the 30 known objects | AlexNet · Known training-set objects | error rate: 99.67 -> 6.7 | Helped |
| Train with adversarial poses from known objects | AlexNet · Held-out objects | error rate: 99.81 -> 89.2 | Helped |
| Test Inception-v3 adversarial poses on other ImageNet classifiers | AlexNet and ResNet-50 · ImageNet AX transfer | misclassification rate: 99.9% and 99.4% | Mixed |
| Transfer Inception-v3 adversarial poses to an MS COCO detector | YOLOv3 · MS COCO AX transfer | misclassification rate: 75.5% | Helped |
| Vary lighting among bright, medium, and dark settings | Inception-v3 pose evaluation · 30-object 6D pose space | accuracy difference: 2.29% | No clear effect |
| Disturb pitch from correctly classified poses | Inception-v3 · Single-parameter sensitivity | failure rate: 83 | Hurt |
| Disturb roll from correctly classified poses | Inception-v3 · Single-parameter sensitivity | failure rate: 81 | Hurt |
| Add 50 rotations per optimization step in DR-G (ablation) | DR-G targeted pose optimization · Targeted adversarial poses | optimization success | Helped |
| Use six camera views instead of single-view optimization (ablation) | DR-G · Bald eagle and tiger cat targets | success rate: 0.44 -> 0.23; 0.15 -> 0.04 | Hurt |
Closest papers by meaning
- Progress and Limitations of Deep Networks to Recognize Objects in Unusual Poses (AAAI 2023 · similarity 0.89)
- Towards Robust 3D Pose Transfer with Adversarial Learning (CVPR 2024 · similarity 0.87)
- Adversarial Attacks Beyond the Image Space (CVPR 2019 · similarity 0.86)
- ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations (ICLR 2023 · similarity 0.85)
- 3D Bounding Box Estimation Using Deep Learning and Geometry (CVPR 2017 · similarity 0.85)
- Source-Free and Image-Only Unsupervised Domain Adaptation for Category Level Object Pose Estimation (ICLR 2024 · 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.