EMBLEM: Enhancing Multi-script Table Detection through Masking
Dhruv Kudale, Udhay Brahmi, Ganesh Ramakrishnan
cs.LG
Quick verdict
Claims masking fine-tuning lifts multi-script table detection by 20.8 F1 on new MANDALA benchmark; real gain vs pretrained, smaller vs matched fine-tuning.
Summary
The paper addresses table detection in multi-script documents, where models trained on English/Chinese data degrade sharply. It contributes (1) MANDALA, a manually annotated benchmark of 2,323 table-containing pages across 18 languages, 15 scripts and 7 domains, intended as a test set, and (2) EMBLEM, a fine-tuning paradigm that masks script-specific strokes via Otsu binarization and contour-based rectangle filling (1% size threshold), letting models pre-trained on English data learn layout-only cues. Fine-tuning YOLO, TATR and DocLayout-YOLO on ~2,000 masked English-dominant pages, EMBLEM-DocL reaches 90.4 F1 at IoU 0.5 on MANDALA, a 20.8-point absolute gain over the pretrained DocL baseline, while staying competitive on ICDAR 2013 (99.0 F1 for TATR), cTDaR, PubTables, IIIT-AR-13K and TNCR. An ablation shows fine-tune-on-masked, infer-on-original works best. Key caveats: the headline gain is against pretrained models; against models fine-tuned on the same unmasked images the margin shrinks (e.g., TATR 86.4 vs 85.7 on MANDALA), the masking threshold was selected on the cTDaR test set, no error bars or significance tests are reported, and the claim of 'consistently' beating all baselines is contradicted by several Table 3 cells. Code and data are released.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Fine-tuning on script-masked images lets TD models pretrained on English/Chinese data generalize to 15 scripts without any multi-script training labels.
- EMBLEM-DocL reaches 90.4 F1 at IoU 0.5 on MANDALA, a 20.8-point absolute gain over its pretrained baseline.
- EMBLEM variants remain competitive on five English-dominant benchmarks (e.g., EMBLEM-TATR 99.0 F1 on ICDAR 2013, EMBLEM-YOLO 98.1 AP50 on PubTables).
- Fine-tuning on masked images while inferring on original images is the most effective configuration (Table 8 ablation).
- MANDALA is positioned as a manually annotated TD benchmark spanning 18 languages, 15 scripts, and 7 domains, unlike existing monolingual/Latin-centric resources (Table 1).
Strengths
New genuinely multi-script test benchmark
MANDALA (Section 3, Table 2) provides 2,323 manually annotated pages across 18 languages/15 scripts/7 domains with case-wise annotation rules (unbordered, rotated, large tables, captions, pseudo-tables) and per-language counts of 93-206 — filling a real gap vs Latin-only benchmarks (Table 1).
Matched fine-tuning baselines reported
Table 3 includes FTN rows (same models fine-tuned on the identical unmasked 2,000/183 split), so the masking effect is isolated rather than hidden; the 4-way train/inference ablation (Table 8) cleanly shows Masked-Original is best (MANDALA 90.4 vs Original-Original 78.8 for DocL).
Breadth of evaluation
Reproducibility support
Weaknesses
Headline gain measured against the wrong baseline
'Consistently outperforms all baselines' contradicted by own Table 3
Hyperparameter tuned on test data
No variance or significance reporting
Internal inconsistency about training scripts
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets Multi-script Table Detection (MTD): locating tables in document images whose scripts differ from the English/Chinese data that dominant detectors were trained on. It makes two coupled contributions: **MANDALA**, a manually annotated test benchmark of 2,323 table-containing pages spanning 18 languages, 15 scripts, and 7 domains (Section 3, Tables 1-2), and **EMBLEM**, a fine-tuning paradigm that masks fine-grained text strokes (Otsu binarization, contour extraction, rectangle fill for boxes below 1% of image dimensions; Algorithm 1) so that models pre-trained on English data learn layout-only cues. The headline claim (Abstract, Section 5.2) is a 20.8-point absolute F1 gain on MANDALA for EMBLEM-DocL (90.4 vs 69.6 at IoU 0.5) using only English-dominant masked fine-tuning data, while staying competitive on five English-heavy benchmarks. Scope is detection only: structure recognition and nested tables are explicitly out of scope (Section 7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the EMBLEM method for table detection?
EMBLEM is a fine-tuning strategy in which table-detection models are trained on images whose text strokes have been masked out — via Otsu binarization and filling small contours (under 1% of image size) with rectangles — so the model learns script-agnostic layout cues like lines and whitespace. At inference the original unmasked image is used. On the multi-script MANDALA benchmark this raised DocLayout-YOLO from 69.6 to 90.4 F1 (IoU 0.5) without any multi-script training labels.
What is the MANDALA dataset?
MANDALA (Multi-script Annotated Documents for Table Detection) is a manually annotated benchmark from IIT Bombay with 2,323 table-containing document pages covering 18 languages written in 15 scripts (e.g., Devanagari, Tamil, Khmer, Perso-Arabic, Cyrillic) across 7 domains such as scientific, administrative, and commercial documents. It is intended primarily as a test benchmark for multi-script table detection and is released with the EMBLEM code.
How much does EMBLEM improve multi-script table detection?
On MANDALA, EMBLEM-DocL scores 90.4 F1 at IoU 0.5 versus 69.6 for the same model without fine-tuning — a 20.8-point absolute gain. Against the same model simply fine-tuned on unmasked images the gain is smaller (for example, 86.4 vs 85.7 F1 for the TATR variant), and on some English benchmarks the plain fine-tuned baselines stay slightly ahead.
Is the EMBLEM code and the MANDALA dataset publicly available?
Yes. The paper states that MANDALA, the code, and the fine-tuned models are released at github.com/IITB-LEAP-OCR/EMBLEM, and the fine-tuning data comes from public Kaggle datasets. Experiments ran on a single NVIDIA A6000, so the pipeline is reproducible for most labs with one GPU.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Fig. 1-4) appear only as captions in the extracted text; the qualitative masking and prediction visuals could not be assessed, though all quantitative claims were verifiable from the tables.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.