Research map / Architectures and efficiency / RNNs and normalization
Bridging the Gap Between End-to-End and Two-Step Text Spotting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Bridging Text Spotting, which connects a frozen well-trained text detector and recognizer via a zero-initialized Bridge plus Adapters, training only the bridge and adapters to get modular two-step systems that match end-to-end accuracy.
Problem
Two-step text spotting is preferred in competitions and practice for its modularity but suffers from error accumulation and sub-optimal performance, while end-to-end spotters sacrifice modularity and require costly full retraining.
Approach
Lock a pretrained detector and recognizer and join them with a zero-initialized Bridge (zero-initialized conv and linear layers around a Transformer encoder) that injects large-receptive-field detection features into the recognizer, plus Adapters inside both modules to learn end-to-end optimization; only the Bridge and Adapters are trained.
Key results
- DG-Bridge Spotter reaches 83.3% end-to-end F on Total-Text (None lexicon), up from 77.5% for the two-step baseline and above ESTextSpotter's 80.8%.
- On CTW1500 the method achieves 69.8% None-lexicon end-to-end accuracy, a 4.9% margin over the previous state of the art (ESTextSpotter, 64.9%).
- Bridge improves the two-step baseline by 4.0% (DG), 2.8% (TM), and 1.6% (BP) on Total-Text, with only a small speed drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| connect DPText-DETR and DiG using Bridging Text Spotting | DPText-DETR + DiG · Total-Text | End-to-end accuracy (None): 77.5 -> 83.3 | Helped |
| connect DPText-DETR and DiG using Bridging Text Spotting | DPText-DETR + DiG · CTW1500 | End-to-end accuracy (None): 64.9 -> 69.8 | Helped |
| connect TESTR detector and DiG using Bridging Text Spotting | TESTR detector + DiG · ICDAR 2015 | End-to-end accuracy (Generic lexicon): 78.1 -> 80.4 | Helped |
| add Bridge (ablation) | DPText-DETR + DiG · Total-Text | End-to-end accuracy: 77.5 -> 81.5 | Helped |
| add detector Adapter (ablation) | DPText-DETR + DiG with Bridge · Total-Text | End-to-end accuracy: 81.5 -> 82.1 | Helped |
| add recognizer Adapter (ablation) | DPText-DETR + DiG with Bridge and detector Adapter · Total-Text | End-to-end accuracy: 82.1 -> 83.3 | Helped |
| add Bridge (ablation) | TESTR detector + MAERec · Total-Text | End-to-end accuracy: 78.0 -> 80.8 | Helped |
| add Bridge (ablation) | DBNet + PARSeq · Total-Text | End-to-end accuracy: 69.1 -> 70.7 | Helped |
| use 1 Transformer layer instead of 0 in Bridge (ablation) | DPText-DETR + DiG · Total-Text | End-to-end accuracy: 82.3 -> 83.3 | Helped |
| use 3 Transformer layers instead of 1 in Bridge (ablation) | DPText-DETR + DiG · Total-Text | End-to-end accuracy: 83.3 -> 82.6 | Hurt |
| use 6 Transformer layers instead of 1 in Bridge (ablation) | DPText-DETR + DiG · Total-Text | End-to-end accuracy: 83.3 -> 82.9 | Hurt |
| use zero-initialized rather than Gaussian-initialized Bridge layers (ablation) | DPText-DETR + DiG · Total-Text | End-to-end accuracy: 78.4 -> 81.5 | Helped |
Surprising findings
- Naive end-to-end training with the same detector and recognizer scored lower (75.6) than the two-step pipeline (77.5) on Total-Text despite comparable detection performance.
- Gaussian-initialized bridge instead of zero-initialized only raised F from 77.5% to 78.4%, while zero-init reached 81.5%.
Limitations
- Gains on the Full lexicon are smaller than on the None lexicon, with results comparable to the baseline there.
- Evaluated only on scene-text benchmarks with detector/recognizer pairs chosen by the authors; multi-task extension left for future work.
Experimental setup
Datasets: Total-Text CTW1500 ICDAR 2015 TextOCR HierText
Benchmarks: Total-Text CTW1500 ICDAR 2015
Metrics: precision recall F-measure end-to-end accuracy with None/Full lexicon Strong/Weak/Generic lexicon accuracy FPS parameters
Baselines: DPText-DETR+DiG TESTR-det+MAERec DB+PARSeq ESTextSpotter DeepSolo SPTS v2 GLASS SRSTS
Builds on: DPText-DETR DiG Adapter (parameter-efficient transfer) zero-initialized ControlNet-style layers TESTR DBNet PARSeq
Closest papers by meaning
- GoMatching: A Simple Baseline for Video Text Spotting via Long and Short Term Matching (NeurIPS 2024 · similarity 0.88)
- SemiETS: Integrating Spatial and Content Consistencies for Semi-Supervised End-to-end Text Spotting (CVPR 2025 · similarity 0.87)
- InstructOCR: Instruction Boosting Scene Text Spotting (AAAI 2025 · similarity 0.84)
- Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective (NeurIPS 2025 · similarity 0.84)
- Arbitrary Reading Order Scene Text Spotter with Local Semantics Guidance (AAAI 2025 · similarity 0.84)
- General Detection-based Text Line Recognition (NeurIPS 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.