Research map / Label-efficient and robust learning / Domain adaptation and generalization
Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net
Xingang Pan, Ping Luo, Jianping Shi, Xiaoou Tang
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 |
|---|---|---|---|
| replace half of BN with IN in shallow residual groups (IBN-a) | ResNet50 · ImageNet val | top-1 error: 24.27 -> 22.54 | Improved |
| add IN after addition in early layers (IBN-b) | ResNet50 · ImageNet val, Monet style | top-1 error: 54.51 -> 50.45 | Improved |
| add IBN-b to ResNet50 under std x1.5 transform | ResNet50 · ImageNet val, std x1.5 | top-1 error: 40.01 -> 23.64 | Improved |
| apply IBN-a to other backbones | DenseNet169, ResNet101, ResNeXt101, SE-ResNet101 · ImageNet val | top-1 error: 0.79%, 1.09%, 0.43%, 0.43% | Improved |
| add IN to residual groups 1-3 vs none (ablation) | IBN-Net50-a · ImageNet val | top-1 error: 24.27 -> 22.54 | Improved |
| add IN also to last residual group (group 4) (ablation) | IBN-Net50-a · ImageNet val | top-1 error: 22.54 -> 22.96 | Hurt |
| raise IN channel ratio from 0.5 to 1 (ablation) | IBN-Net50-a · ImageNet val | top-1 error: 22.54 -> 23.44 | Hurt |
| replace all BN with IN (ablation) | ResNet50 · ImageNet val | top-1 error: 24.27 -> 28.56 | Hurt |
| add additional IBN layers (a x2) (ablation) | IBN-Net50-a · ImageNet val | top-1 error: 22.54 -> 22.81 | No clear effect |
| combine IBN-a with IBN-d (ablation) | IBN-Net50-d · ImageNet val | top-1 error: 22.86 -> 22.89 | No clear effect |
| IBN-a trained on Cityscapes, tested on GTA5 | ResNet50 · Cityscapes to GTA5 | mIoU: 29.4 -> 32.5 | Improved |
| IBN-b trained on Cityscapes, tested on GTA5 | ResNet50 · Cityscapes to GTA5 | mIoU: 29.4 -> 37.9 | Improved |
Closest papers by meaning
- The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes (NeurIPS 2022 · similarity 0.87)
- Ubernet: Training a Universal Convolutional Neural Network for Low-, Mid-, and High-Level Vision Using Diverse Datasets and Limited Memory (CVPR 2017 · similarity 0.87)
- Deep SimNets (CVPR 2016 · similarity 0.87)
- XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016 · similarity 0.87)
- Refining Architectures of Deep Convolutional Neural Networks (CVPR 2016 · similarity 0.87)
- Interpolating Convolutional Neural Networks Using Batch Normalization (ECCV 2018 · 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.
