Research map / Label-efficient and robust learning / Recognition and long-tail
TDeLTA: A Light-Weight and Robust Table Detection Method Based on Learning Text Arrangement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TDeLTA, a 3.1M-parameter table detection model that classifies text blocks (from PDF parsers or OCR) into four semantic roles using a BiLSTM plus multi-head attention, achieving strong zero-shot cross-domain robustness.
Problem
Image-based table detectors overfit to table visual styles and degrade sharply on out-of-distribution tables from other domains, while also being large and slow.
Approach
Treat table detection as classifying text blocks by semantic role (row header, column header, content cell, outside text) using only text-block position features, modeled with a BiLSTM encoder and multi-head attention, then derive table boundaries via post-processing.
Key results
- On PubTables-1M, TDeLTA reaches average F1 98.70 (IoU 0.5-0.95) with 3.1M parameters, competitive with Table-DETR (97.82) and CascadeTabNet (99.41) at 28.9M and 82.6M parameters.
- Under 0-shot transfer to FinTabNet with PDF-parser text blocks, TDeLTA gets average F1 80.46 versus 73.45 for DiT, a gap of about 7 points.
- With PaddleOCR-extracted text blocks on FinTabNet 0-shot, TDeLTA gets average F1 78.30, still above all image-based baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BiLSTM instead of Transformer, LSTM, or BiGRU as the sequence encoder (ablation) | TDeLTA · PubTables-1M | Avg. F1 (IoU@0.5-0.95): 98.45 -> 98.70; 96.11 -> 98.70; 97.47 -> 98.70 | Helped |
| Use BiLSTM instead of Transformer, LSTM, or BiGRU as the sequence encoder (ablation) | TDeLTA · FinTabNet test, 0-shot, PDF parser | Avg. F1 (IoU@0.5-0.95): 71.99 -> 80.46; 58.60 -> 80.46; 71.18 -> 80.46 | Helped |
| Use BiLSTM instead of Transformer, LSTM, or BiGRU as the sequence encoder (ablation) | TDeLTA · FinTabNet test, 0-shot, open-source OCR tool | Avg. F1 (IoU@0.5-0.95): 70.49 -> 78.30; 63.10 -> 78.30; 69.24 -> 78.30 | Helped |
| Increase BiLSTM layers from 2 to 8 (ablation) | TDeLTA · PubTables-1M | AP: 97.23 -> 98.02 | Helped |
| Increase BiLSTM layers from 2 to 8 (ablation) | TDeLTA · FinTabNet | AP: 67.32 -> 74.89 | Helped |
| Increase BiLSTM layers from 8 to 12 (ablation) | TDeLTA · PubTables-1M | AP: 98.02 -> 97.91 | Hurt |
| Increase BiLSTM layers from 8 to 12 (ablation) | TDeLTA · FinTabNet | AP: 74.89 -> 76.08 | Helped |
| Increase BiLSTM hidden size from 16 to 128 (ablation) | TDeLTA · PubTables-1M | AP: 88.54 -> 98.02 | Helped |
| Increase BiLSTM hidden size from 16 to 128 (ablation) | TDeLTA · FinTabNet | AP: 41.66 -> 74.89 | Helped |
| Use OCR-extracted text blocks instead of PDF-parser text blocks | TDeLTA · FinTabNet test, 0-shot | Avg. F1 (IoU@0.5-0.95): 80.46 -> 78.30 | Hurt |
Surprising findings
- A 0.119M-parameter variant (hidden size 32) already outperforms all baselines on FinTabNet 0-shot in the ablation.
- On PubTables-1M, TDeLTA trails the strongest in-domain baselines (CascadeTabNet, Table-DETR) slightly, while the image-based baselines collapse out of domain.
Limitations
- Errors on adjacent tables with identical formats, standalone short sentences above tables, and tables with long text content.
- Relies on text-block positions from a PDF parser or OCR, so text extraction quality affects results.
- Trails the strongest in-domain baselines slightly on PubTables-1M.
Experimental setup
Datasets: PubTables-1M FinTabNet
Benchmarks: PubTables-1M FinTabNet
Metrics: precision recall F1 at IoU@0.5 F1 at IoU@0.6 average F1 IoU@0.5-0.95 parameter count
Baselines: Table-DETR CascadeTabNet DiT YOLOv7 LayoutLMv3 StructTextV2
Builds on: BiLSTM multi-head attention Cascade R-CNN
Closest papers by meaning
- TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition (CVPR 2026 · similarity 0.87)
- Grab What You Need: Rethinking Complex Table Structure Recognition with Flexible Components Deliberation (AAAI 2024 · similarity 0.85)
- TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.83)
- HeGTa: Leveraging Heterogeneous Graph-enhanced Large Language Models for Few-shot Complex Table Understanding (AAAI 2025 · similarity 0.83)
- Large Scale Transfer Learning for Tabular Data via Language Modeling (NeurIPS 2024 · similarity 0.83)
- TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.