Research map / Image generation / Editing and style transfer
SimAN: Exploring Self-Supervised Representation Learning of Scene Text via Similarity-Aware Normalization
Canjie Luo, Lianwen Jin, Jingdong Chen
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 similarity-based local style alignment (SimAN) instead of global-statistics AdaIN (ablation) | ResNet-FCN-Att. probe · IIIT5K | accuracy: 9.7 -> 22.2 | Improved |
| Use similarity-based local style alignment (SimAN) instead of global-statistics AdaIN (ablation) | ResNet-FCN-Att. probe · IC03 | accuracy: 9.1 -> 22.3 | Improved |
| Use similarity-based local style alignment (SimAN) instead of global-statistics AdaIN (ablation) | ResNet-FCN-Att. probe · IC13 | accuracy: 11.1 -> 24.1 | Improved |
| Add adversarial training (ablation) | SimAN representation learner · Scene-text probe (IIIT5K, IC03, IC13 average) | accuracy (%): 68.8 -> 69.0 | Improved |
| Add adversarial training (ablation) | SimAN image generator · Generated image distribution | FID: 24.0 -> 23.2 | Improved |
| Add self-supervised pretraining on unlabeled SynthText | SimAN recognizer with 10K labeled SynthText · IIIT5K | word accuracy (%): 35.0 -> 41.1 | Improved |
| Add self-supervised pretraining on unlabeled SynthText | SimAN recognizer with 10K labeled SynthText · IC03 | word accuracy (%): 37.6 -> 42.9 | Improved |
| Add self-supervised pretraining on unlabeled SynthText | SimAN recognizer with 10K labeled SynthText · IC13 | word accuracy (%): 38.6 -> 43.9 | Improved |
| Pretrain the encoder on Real-300K rather than SynthText | SimAN frozen-backbone probe with SynthText-trained probe · IIIT5K | accuracy: 60.8 -> 68.9 | Improved |
| Add a sequence model (2*RNN) to the self-supervised encoder (ablation) | SimAN CTC probe · IIIT5K | accuracy: 1.5 -> 0.0 | Hurt |
| Add a sequence model (2*RNN) to the self-supervised encoder (ablation) | SimAN attention probe · IIIT5K | accuracy: 22.2 -> 6.4 | Hurt |
| Use two BiLSTMs rather than one in the probe (ablation) | SimAN CTC probe · IIIT5K | accuracy: 57.4 -> 60.8 | Improved |
Closest papers by meaning
- Perceiving Stroke-Semantic Context: Hierarchical Contrastive Learning for Robust Scene Text Recognition (AAAI 2022 · similarity 0.86)
- Primitive Representation Learning for Scene Text Recognition (CVPR 2021 · similarity 0.86)
- SEE: Towards Semi-Supervised End-to-End Scene Text Recognition (AAAI 2018 · similarity 0.85)
- Boosting Semi-Supervised Scene Text Recognition via Viewing and Summarizing (NeurIPS 2024 · similarity 0.85)
- OTE: Exploring Accurate Scene Text Recognition Using One Token (CVPR 2024 · similarity 0.84)
- Linguistics-aware Masked Image Modeling for Self-supervised Scene Text Recognition (CVPR 2025 · similarity 0.83)
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.
