Research map / Detection and segmentation / Open-set and few-shot detection
Progress and Limitations of Deep Networks to Recognize Objects in Unusual Poses
Read the paperCheck 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 |
|---|---|---|---|
| Evaluate objects in unusual poses rather than upright poses | 38-network collection · ObjectPose vs. ObjectPose +-10 | top-1 accuracy: average accuracy drop of 29.5\% | Hurt |
| Train on a very large dataset and use a large-capacity model | Image classification networks · ObjectPose | top-1 accuracy: 14.5\% accuracy drop | Helped |
| Increase model capacity from EfficientNet-B7 to EfficientNet-L2 while using Noisy Student on JFT-300M | Noisy Student EfficientNet · ObjectPose | accuracy | Helped |
| Use CLIP prompts describing an unusual pose (ablation) | CLIP · ObjectPose | accuracy | Helped |
| Combine rotations along ROLL, PITCH, and YAW instead of rotating along one axis | All networks · ObjectPose, three-axis rotation | accuracy: 6\%-16.5\% | Hurt |
| Combine three-axis rotations with scaling | All networks · ObjectPose, combined transformations | accuracy: 24.5\%-78.2\% compared to usual poses | Hurt |
| Use a smaller object scale | Most networks · ObjectScale | accuracy | Hurt |
| Use in-plane rotations rather than out-of-plane rotations | Most networks · In-plane vs. out-of-plane rotation | accuracy | Helped |
| Rotate the background together with the object instead of keeping the background upright | Best networks, including Noisy Student EfficientNets and SWAG · Image rotation vs. in-plane rotation | accuracy | Helped |
| Rotate the background together with the object instead of keeping the background upright | Weaker networks · Image rotation vs. in-plane rotation | accuracy | Hurt |
| Remove natural backgrounds (ablation) | Best networks · ObjectPose | accuracy | Helped |
| Rotate the background alone (ablation) | Weaker networks · Background Rotation | accuracy | Hurt |
Closest papers by meaning
- Contemplating Real-World Object Classification (ICLR 2021 · similarity 0.91)
- Strike (With) a Pose: Neural Networks Are Easily Fooled by Strange Poses of Familiar Objects (CVPR 2019 · similarity 0.89)
- Generalisation in humans and deep neural networks (NeurIPS 2018 · similarity 0.88)
- Harmonizing the object recognition strategies of deep neural networks with humans (NeurIPS 2022 · similarity 0.88)
- Harnessing Edge Information for Improved Robustness in Vision Transformers (AAAI 2024 · similarity 0.86)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.86)
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.