- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents a domain-bootstrapped OCR pipeline for recognizing Traditional Chinese Painting Inscriptions (TCPIs) under limited high-quality supervision. It first constructs a small seed dataset through a machine–expert annotation protocol that combines candidate transcriptions from three pretrained OCR models with expert verification. A general-purpose OCR model is then fine-tuned on the seed dataset to obtain a TCPI-adapted model, which generates pseudo-labels for a larger unlabeled collection. After filtering these pseudo-labels, the resulting expansion dataset is used to train the TCPI-bootstrapped model for domain-specific inscription recognition.
Critical strengths
- The paper addresses a meaningful and challenging cultural heritage application. TCPI recognition involves expressive calligraphy, irregular layouts, background interference, and image degradation, making it a practically valuable setting for studying low-resource OCR adaptation.
- The pipeline provides a practical approach to constructing domain-specific supervision. Combining multi-model candidate generation, expert verification, TCPI adaptation, and pseudo-label expansion offers a coherent workflow for using a small verified seed dataset to support larger-scale training.
- The reported validation results show clear benefits from domain adaptation. On the 94-sample seed validation set, character-level F1 improves from 0.825 for the base olmOCR-2 model to 0.861 for TCPI-adapted and 0.868 for TCPI-bootstrapped. These results provide evidence that domain-specific supervision benefits TCPI recognition.
- The ablation studies provide useful empirical observations about training choices. The paper examines pseudo-labeled training scale, augmentation schemes, and FPFT/LoRA/RL strategies, with additional breakdowns by clarity, calligraphy, interference, and text length. These analyses reveal potentially useful domain-specific considerations, such as the adverse effect of rotation augmentation under the evaluated settings.
- The expert audit improves transparency regarding pseudo-label quality. The paper independently evaluates pseudo-labels against expert transcriptions on 463 expansion samples and reports both aggregate and length-specific results. This makes the remaining label noise explicit and provides useful evidence about the supervision used for bootstrapped training.
Critical weaknesses
- The paper states that approximately 12,000 paintings were used in the data construction pipeline, but it does not identify the public datasets, institutions, or websites from which the images were obtained. The collection procedure and applicable usage conditions are also unspecified. This leaves potential copyright and licensing concerns unresolved.
- The paper presents a complete and reproducible TCPI OCR pipeline, but it does not explicitly identify the final training configuration used for the TCPI-bootstrapped model in the main experiments. The Overall metrics in Table 2 match those of the “FPFT + W/o Rotate” configuration in Tables 3–5, leaving readers to infer the configuration indirectly.
- The paper uses a fine-tuned DINO-YOLO inscription detector to localize inscriptions within paintings, but it does not clarify the specific DINO and YOLO variants involved, how the components are combined, or which parameters are fine-tuned. It also does not report detector performance, such as mAP or a clearly defined [email protected]. Furthermore, the source and size of the detector fine-tuning dataset, preprocessing procedures, bounding-box annotation protocol, and training/validation split are not specified.
- Important training configurations are insufficiently reported across the three training strategies and experiments. Missing details include learning rates, optimizers, batch sizes, and training epochs; LoRA target layers, , and ; and the RL optimization algorithm, number of samples per input, sampling temperature, and regularization terms and coefficients. If a KL penalty is used, its formulation and coefficient should also be reported. These omissions hinder reproducibility and make it difficult to determine whether, and to what extent, the performance differences in Tables 2 and 3 are attributable to differences in hyperparameter configurations.
- The ablation studies in Tables 3–5 are primarily evaluated on the Overall set. Although this set can assess in-domain knowledge retention, approximately 90% of its samples were used during the initial adaptation stage, so it cannot provide an independent assessment of generalization. Even if the second stage does not initialize from the TCPI-adapted model, supervision from the seed training set can still influence the final model through pseudo-label generation. Consequently, the current ablations do not adequately establish the contribution of each component to generalization on unseen samples.
Questions for authors
- Could the authors provide the data sources, collection procedures, copyright or licensing information, and redistribution conditions discussed in Weakness 1?
- Regarding Weakness 2, could the authors explicitly specify the final training strategy, augmentation scheme, and model-selection criteria? Do the Validation and Overall results in Table 2 come from the same checkpoint?
- Could the authors provide the detector architecture, training data, fine-tuning procedure, and evaluation details discussed in Weakness 3?
- Regarding Weakness 4, could the authors report the important hyperparameter configurations for all three training strategies across the experiments?
- Regarding Weakness 5, could the authors report the corresponding ablation results on a completely held-out validation set or an independent test set?
- Could the authors provide the complete computation of the rule-based score used for pseudo-label filtering? The paper mentions special symbols and repetitions, but does not specify their precise definitions, penalty weights, normalization procedures, or handling of anomalous outputs. A formula or pseudocode would help make this step reproducible.
- In the second-stage training that produces the TCPI-bootstrapped model , are the parameters initialized from the TCPI-adapted model , the base olmOCR-2 model , or another checkpoint? Do FPFT, LoRA, and RL use the same initialization? Does this stage use only the pseudo-labeled expansion dataset, or is the seed training dataset also included?
- In Table 2, TCPI-bootstrapped has an EM score that is 2.1 percentage points lower than TCPI-adapted on Validation, despite improvements in F1 and NES. Could the authors analyze possible reasons for this decrease after training on the expansion dataset, including whether pseudo-label noise, the training objective, or other configuration changes contribute to it?
- The paper reports that RL underperforms FPFT and LoRA overall and discusses its localized advantage on long inscriptions, but does not sufficiently analyze the reasons for its lower overall performance. Could the authors discuss whether the mismatch between the reward objective and evaluation metrics, pseudo-label noise, or training configurations may contribute to this result?
- The complete training experiments are conducted only with olmOCR-2, so the applicability of the pipeline to other OCR backbones remains unclear. Could the authors evaluate one or two additional trainable OCR models with different architectures or pretraining backgrounds to determine whether they also benefit from this pipeline?