List your company
ResearchFormerly WizWand

Handwriting Recognition on IAM

2.38CERDTrOCR
2.21923.30464.395.4754Apr 15, 2021Feb 10, 2022Dec 9, 2022Oct 7, 2023Aug 3, 2024Jun 1, 2025Mar 30, 2026
Updated 7mo ago

Evaluation results

MethodLinks
DTrOCRTraining Data=Synthetic + IAM, External LM=No2023.082.38—
DTrOCRArchitecture=Decoder-Only Transform., Params=300M, Training Data=2B+ Synthetic Lines, Adaptation Cost=>500 GPU hrs2026.032.38—
S-Attn/CTC + LMRect.=No, Encoder/Decoder=S-Attn/CTC + LM, Train Dataset=Internal + Public2021.042.75—
Diaz et al.Architecture=S-Attn/CTC, Training Data=Internal + IAM, External LM=Yes2021.092.75—
Diaz et al.Training Data=Internal + IAM, External LM=Yes2023.082.75—
TrOCR_LARGEArchitecture=Transformer, Training Data=Synthetic + IAM, External LM=No2021.092.89—
TrOCRLARGETraining Data=Synthetic + IAM, External LM=No2023.082.89—
TrOCR-LargeArchitecture=End-to-End Transform., Params=558M, Training Data=684M Synthetic Lines, Adaptation Cost=300+ GPU hrs2026.032.89—
TransformerRect.=No, Encoder/Decoder=Transformer, Train Dataset=Internal + Public2021.042.96—
Diaz et al.Architecture=Transformer w/ CNN, Training Data=Internal + IAM, External LM=No2021.092.96—
S-Attn/CTC + LMRect.=No, Encoder/Decoder=S-Attn/CTC + LM, Train Dataset=Internal2021.043.15—
Bluche and MessinaRect.=No, Encoder/Decoder=GCRNN/CTC, Train Dataset=IAM (50k lexicon)2021.043.2—
Bluche and MessinaArchitecture=GCRNN/CTC, Training Data=Synthetic + IAM, External LM=Yes2021.093.2—
Bluche et al.Training Data=Synthetic + IAM, External LM=Yes2023.083.2—
TrOCR_BASEArchitecture=Transformer, Training Data=Synthetic + IAM, External LM=No2021.093.42—
TrOCR-BaseArchitecture=End-to-End Transform., Params=334M, Training Data=684M Synthetic Lines, Adaptation Cost=200+ GPU hrs2026.033.42—
S-Attn/CTCRect.=No, Encoder/Decoder=S-Attn/CTC, Train Dataset=Internal + Public2021.043.53—
Diaz et al.Architecture=S-Attn/CTC, Training Data=Internal + IAM, External LM=No2021.093.53—
TransformerRect.=No, Encoder/Decoder=Transformer, Train Dataset=Internal2021.043.99—
transformerLevel=line2025.043.99—
TrOCR_SMALLArchitecture=Transformer, Training Data=Synthetic + IAM, External LM=No2021.094.22—
DANIELLevel=page2025.044.38—
MT-DANLevel=page2025.044.51—
DANLevel=page2025.044.54—
S-Attn/CTCRect.=No, Encoder/Decoder=S-Attn/CTC, Train Dataset=Internal2021.044.62—
Kang et al.Rect.=No, Encoder/Decoder=Transformer, Train Dataset=IAM2021.044.67—
Kang et al.Architecture=Transformer w/ CNN, Training Data=Synthetic + IAM, External LM=No2021.094.67—
Kang et al.Training Data=Synthetic + IAM, External LM=No2023.084.67—
Meta-DANLevel=page2025.044.79—
Michael et al.Rect.=No, Encoder/Decoder=LSTM/LSTM w/Attn, Train Dataset=IAM2021.044.87—
Michael et al.Architecture=LSTM/LSTM w/Attn, Training Data=IAM, External LM=No2021.094.87—
Michael et al.Training Data=IAM, External LM=No2023.084.87—
BART-BaseArchitecture=Linguistic Adapter, Params=140M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 0.5 GPU hrs2026.035.18—
T5-BaseArchitecture=Linguistic Adapter, Params=220M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 0.5 GPU hrs2026.035.4—
Detector BaselineArchitecture=Object Detector Only, Params=40M, Training Data=1M Synthetic, Adaptation Cost=3 GPU hrs2026.036.09—
ByT5-BaseArchitecture=Linguistic Adapter, Params=582M, Training Data=1.5M Synthetic, Adaptation Cost=3 + 1.0 GPU hrs2026.036.35—
Wang et al.Rect.=No, Encoder/Decoder=FCN/GRU, Train Dataset=MJ + ST2021.046.4—
Wang et al.Architecture=FCN/GRU, Training Data=IAM, External LM=No2021.096.4—
Wang et al.Training Data=IAM, External LM=No2023.086.4—
ABINetTraining-free evaluation=true2023.11—61.57
ASTERTraining-free evaluation=true2023.11—52.5
E2STR-baseTraining-free evaluation=true, In-Context Learning=false2023.11—69.51
E2STR-ICLTraining-free evaluation=true, In-Context Learning=true2023.11—74.1
MAERecTraining-free evaluation=true2023.11—70.27
NRTRTraining-free evaluation=true2023.11—59.53
SARTraining-free evaluation=true2023.11—56.63
SATRNTraining-free evaluation=true2023.11—59.47