acceptodds
Withdrawn from ICLR 2027

Hand-Punched Braille Document Recognition: A New Benchmark Dataset and a Robust Detection Transformer

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Optical Braille recognition plays an important role in digitizing Braille documents and supporting accessible learning for blind and visually impaired learners. However, existing page-level Braille benchmarks provide limited coverage of smartphone-captured hand-punched documents with Braille dot appearance variation, which normally exhibit illumination variation and geometric distortion. To address this data gap, we introduce BrailleStuDoc, a benchmark dataset of student-hand-punched Braille documents captured with smartphones. It contains 331 images from distinct physical pages and 149,720 cell-level annotations covering all 63 non-empty six-dot patterns. To jointly address degraded local appearance cues and geometric variations in smartphone-captured hand-punched Braille documents, we propose BRIDGE, an illumination- and geometry-aware Braille cell detection Transformer. BRIDGE introduces Micro-Surface Feature Enhancement (MSFE) to strengthen local appearance representation under illumination and dot appearance variations, and Geometry-Aware Scale Modeling (GASM) to model the local geometric regularity of neighboring Braille cells under geometric distortion. Extensive experiments demonstrate the complementary value of BrailleStuDoc to existing public datasets and the effectiveness of BRIDGE, which achieves the best overall recognition performance among the compared methods.

Primary area:
datasets and benchmarks
Keywords:
Braille Document Recognition, Benchmark Dataset, Detection Transformer

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1Y2SOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a database "BrailleStuDoc" which contains 331 smartphone photos of student hand-punched Chinese Braille pages with 149,720 cell-level boxes over the 63 dot patterns. It also proposes BRIDGE, which adds two modules to RT-DETRv2. MSFE adds an illumination- and reliability-biased attention in the encoder, and GASM refines queries and box scales from row/column neighbours in the last three decoder layers. It reports 99.50 F1 on BrailleStuDoc and the best F1 on DSBI, Angelina and BrailleStuDoc under joint training.

Critical strengths
  1. The database presents an authentic and fully hand-punched student work under uncontrolled phone capture for an accessibility application.
  2. The annotation protocol includes second-annotator review, expert adjudication and tactile checks on the physical page.
Critical weaknesses
  1. The evidence for the method is limited.
  2. The robustness claim is not supported where it matters. Since 266 of 331 images show illumination variation, the remaining test cells would need to be essentially error-free. The paper should give test-set image and cell counts per subset.
  3. FortThe ablation study, nothing isolates the three MSFE cues, GASM’s row/column aggregation versus its scale refinement, the choice of decoder layers, or any hyperparameter.
  4. Parameters and FLOPs are not reported.
  5. No visualization shows that the learned maps correspond to illumination or reliability. They are unsupervised and named by intent.
  6. The baselines should include more recent detector other than RT-DETRv2. Simple alternatives should be tested, such as dewarping or shadow-removal preprocessing or grid-fitting post-processing.
  7. As a benchmark, the dataset is small and comes from one school. The number of writers is not reported, and the split is by page, not by writer. In addition, there is no text transcription, so “Chinese” is incidental at the cell level.

The contribution is a small domain dataset plus task-specific modules on an existing detector. The learning-methods novelty is limited.

Questions for authors
  1. What IoU threshold and matching rule define a correct cell, and how are the excluded “unresolved” instances treated at evaluation?
  2. Who labelled the challenge conditions, and with what criteria?
  3. Was consent obtained from the students or guardians, beyond school authorization? Was there ethics review?
  4. What are the latency hardware and input resolution, and were all baselines run at the same resolution?
Reviewer_kCGhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies optical recognition of smartphone-captured hand-punched Braille documents. The authors introduce BrailleStuDoc, a dataset containing 331 images from distinct physical pages and 149,720 cell-level annotations covering all 63 non-empty six-dot patterns. To address illumination and local appearance variation as well as geometric distortion, the paper proposes BRIDGE, an RT-DETRv2-based detector augmented with Micro-Surface Feature Enhancement (MSFE) and Geometry-Aware Scale Modeling (GASM). MSFE is designed to enhance local Braille appearance cues under illumination and dot-appearance variations, while GASM exploits local row-column organization and scale consistency among neighboring Braille cells. Experiments on BrailleStuDoc, DSBI, and Angelina show that BRIDGE improves over the compared baselines and achieves the best reported F1 scores under the evaluated settings.

Critical strengths
  1. A meaningful and carefully constructed real-world benchmark. BrailleStuDoc fills a clear gap in existing page-level Braille benchmarks by focusing specifically on smartphone-captured, student-produced hand-punched documents. The dataset contains 331 distinct physical pages and 149,720 cell-level annotations, and the annotation procedure appears careful: ambiguous patterns can be checked against the original physical document and are further reviewed by multiple annotators and a Braille expert. I consider this dataset contribution practically meaningful and potentially useful to the OBR community.

  2. Reasonable task-specific modeling with consistent empirical gains. The proposed MSFE and GASM modules are well motivated by characteristics of the task. MSFE targets weak local appearance cues affected by illumination and punching variations, while GASM exploits the locally regular row-column structure and scale consistency of Braille cells. Both modules improve over the RT-DETRv2 baseline individually, and their combination gives the best reported performance.

Critical weaknesses

** The experimental evidence does not sufficiently validate the claimed roles of MSFE and GASM.** The current ablation shows that adding MSFE and/or GASM improves the overall F1 score, but it does not establish that these gains specifically come from overcoming illumination/appearance variation and geometric distortion, respectively. In particular, the paper does not report how the baseline, +MSFE, +GASM, and full model perform separately on the illumination and geometric-distortion subsets. Therefore, alternative explanations such as additional feature processing, high-frequency enhancement, neighborhood aggregation, or scale smoothing cannot be ruled out. Since the illumination-aware and geometry-aware motivations are central to the proposed method, I believe more targeted ablations are needed to substantiate these claims. This concern appears potentially addressable through targeted ablations during rebuttal.

Questions for authors

Q1. Could the authors provide targeted ablations on the illumination-variation and geometric-distortion subsets to show whether MSFE specifically improves robustness to illumination/appearance variation and GASM specifically improves robustness to geometric distortion? A convincing result along this direction could materially improve my assessment of the paper.

Q2. How many students contributed the 331 pages, and are the train/validation/test splits student-disjoint? Since hand-punched Braille may contain student-specific patterns, this clarification is important for assessing generalization to unseen users.

Reviewer_op1HOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces BrailleStuDoc, a new dataset for recognizing hand-punched Braille documents captured by smartphones. It contains 331 images and about 150,000 annotated Braille cells, covering all 63 non-empty six-dot patterns.

The authors also propose BRIDGE, a detection model based on RT-DETRv2. It adds two modules: MSFE to improve visual feature extraction under different lighting conditions, and GASM to model the geometric relationships between neighboring Braille cells. Experiments on BrailleStuDoc and two existing datasets show that BRIDGE achieves an F1 score of 99.50% on BrailleStuDoc, compared with 98.23% for the baseline model. Ablation studies also show that both proposed modules contribute to the improvement.

Critical strengths
  1. BrailleStuDoc contains 331 smartphone-captured pages with 149,720 annotated Braille cells. It covers real-world challenges such as uneven lighting, geometric distortion, and irregular Braille dots. This dataset can help researchers study Braille recognition under more practical conditions.

  2. BRIDGE achieves an F1 score of 99.50%, compared with 98.23% for RT-DETRv2. Ablation studies show that both proposed modules contribute to the improvement. Experiments on two existing datasets also provide evidence of the method's effectiveness beyond the new dataset.

Critical weaknesses
  1. The dataset contains only 331 pages collected from one school, with around 33 pages in the test set. This may limit the reliability of the results and make it difficult to know whether the method generalizes to other schools and capture conditions.

  2. The dataset is highly imbalanced, with only 11 examples of the rarest class. The paper mainly reports overall F1 scores, so it is unclear how well BRIDGE recognizes rare patterns. This is important for a reliable Braille recognition system.

Questions for authors
  1. The dataset is highly imbalanced, with only 11 examples of the rarest Braille pattern. Could the authors report per-class performance, especially for rare patterns, to show whether BRIDGE can recognize them reliably?

  2. Since all images were collected from one school, could the authors clarify whether the training and test sets differ in students, writing styles, and capture conditions? This would help assess how well the method generalizes to new data.

  3. Evaluating the method on a larger dataset collected from multiple schools would provide stronger evidence of its generalization ability.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.