Line-level recognition on Antiqua (test)
1.83CERMamba-AR
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Mamba-ARModel Category=Mamba-based (ours), Latency=53.3, Throughput=18.7, Inf Mem=2898, Train Mem=8.062026.04 | 1.83 | 3.84 | |
| DANModel Category=Transformer baselines, Latency=156.5, Throughput=6.4, Inf Mem=706, Train Mem=6.742026.04 | 1.83 | 3.61 | |
| VANModel Category=Transformer baselines, Latency=6.19, Throughput=161.5, Inf Mem=1379, Train Mem=14.142026.04 | 1.85 | 3.39 | |
| Mamba-CTCModel Category=Mamba-based (ours), Latency=7.74, Throughput=129.2, Inf Mem=543, Train Mem=7.172026.04 | 2.18 | 5.16 | |
| DANIELModel Category=Transformer baselines, Latency=23.8, Throughput=41.6, Inf Mem=2986, Train Mem=10.802026.04 | 2.37 | 4.99 | |
| Mamba-NARModel Category=Mamba-based (ours), Latency=13.6, Throughput=73.5, Inf Mem=960, Train Mem=6.312026.04 | 2.66 | 8.26 | |
| TrOCRModel Category=Off-the-shelf systems, Evaluation Protocol=fine-tuned, Latency=284.3, Throughput=3.5, Inf Mem=1472, Train Mem=27.702026.04 | 3.62 | 8.95 | |
| Tesseract v5Model Category=Off-the-shelf systems, Latency=97.9, Throughput=10.22026.04 | 4.79 | 16.06 | |
| GeminiModel Category=Off-the-shelf systems, Latency=∼1000, Throughput=∼12026.04 | 7.23 | 19.11 | |
| PERO-OCRModel Category=Off-the-shelf systems, Latency=30.9, Throughput=32.42026.04 | 13.5 | 26.19 | |
| TrOCRModel Category=Off-the-shelf systems, Evaluation Protocol=zero-shot, Latency=69.5, Throughput=14.4, Inf Mem=13892026.04 | 15.46 | 60.97 |