Handwriting Recognition on CVL Modern Clean Handwriting (test)
82.3Word AccuracyMGP-STR
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MGP-STRArchitecture=Multi-Granularity ViT, Params=148M, Training Data=100M+ Synthetic, Adaptation Cost=100+ GPU hrs2026.03 | 82.3 | — | |
| T5-BaseArchitecture=Linguistic Adapter, Params=220M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 0.5 GPU hrs2026.03 | 78.1 | 1.9 | |
| BART-BaseArchitecture=Linguistic Adapter, Params=140M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 0.5 GPU hrs2026.03 | 77.82 | 1.95 | |
| ByT5-BaseArchitecture=Linguistic Adapter, Params=582M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 1.0 GPU hrs2026.03 | 76.45 | 1.98 | |
| TrOCR-BaseArchitecture=End-to-End Transform., Params=334M, Training Data=684M Synth Lines, Adaptation Cost=200+ GPU hrs2026.03 | 74.5 | 3.42 | |
| Detector BaselineArchitecture=Object Detector Only, Params=40M, Training Data=1M Synthetic, Adaptation Cost=3 GPU hrs2026.03 | 72.84 | 3.71 |