Research map / Label-efficient and robust learning / Recognition and long-tail
Rotation Regularization without Rotation
Takumi Kobayashi
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 statistical rotation instead of geometric rotation matrices (ablation) | rotation regularization · ImageNet-LT | error rate (%): 56.68 -> 56.40 | Improved |
| use correlated random unit vector instead of i.i.d. Gaussian noise (ablation) | StatRot · ImageNet-LT | error rate (%): 56.40 -> 56.41 | No clear effect |
| add the random rotation noise term on top of plain logit scaling (ablation) | StatRot · ImageNet-LT | error rate (%): 58.25 -> 56.40 | Improved |
| use angular cone instead of Euclidean margin ball (ablation) | StatRot · ImageNet-LT | error rate (%): 56.94 -> 56.40 | Improved |
| make noise asymmetric (target logit only, -|epsilon|) (ablation) | StatRot · ImageNet-LT | error rate (%): 56.40 -> 56.45 | No clear effect |
| combine StatRot with DropOut (ablation) | StatRot · ImageNet-LT | error rate (%): 56.40 -> 56.12 | Improved |
| use theoretical D = min(d,C) = 512 instead of smaller D = 128 (ablation) | StatRot · ImageNet-LT | error rate (%) | Improved |
| set rotation angle to pi/4 rather than larger angles (ablation) | StatRot · ImageNet-LT | error rate (%) | Improved |
| add StatRot to softmax training | ResNet10 · ImageNet-LT | error rate (%): 58.91 -> 56.40 | Improved |
| add StatRot to softmax training | ResNet50 · iNat2018 | error rate (%): 32.82 -> 30.39 | Improved |
| add StatRot to softmax training | ResNet10 · Places-LT | error rate (%): 72.99 -> 71.85 | Improved |
| add StatRot to softmax training | ResNet152 pretrained · Places-LT | error rate (%): 61.14 -> 59.93 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ArcFace — ArcFace: Additive Angular Margin Loss for Deep Face Recognition (CVPR 2019) · better on 1 of 1 shared result
- CosFace — CosFace: Large Margin Cosine Loss for Deep Face Recognition (CVPR 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Representation Learning by Rotation Feature Decoupling (CVPR 2019 · similarity 0.85)
- Feature Variance Regularization: A Simple Way to Improve the Generalizability of Neural Networks (AAAI 2020 · similarity 0.85)
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.85)
- Learning with 3D rotations, a hitchhiker's guide to SO(3) (ICML 2024 · similarity 0.85)
- Regularization on Spatio-Temporally Smoothed Feature for Action Recognition (CVPR 2020 · similarity 0.84)
- T-vMF Similarity for Regularizing Intra-Class Feature Distribution (CVPR 2021 · 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.
