Research map / Architectures and efficiency / RNNs and normalization
Grab What You Need: Rethinking Complex Table Structure Recognition with Flexible Components Deliberation
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 GrabTab, a table structure recognition model that deliberates over multiple table components (Bézier separators, element and row/column relation tokens) through a Components Deliberator to predict cell structure without rule-based post-processing.
Problem
Table structure recognition for complex tables, including camera-captured images with geometric distortion and irregular inner structure, where single-component methods (boundary- or relation-based) fail and rely on heuristic post-processing.
Approach
Treat table separators as the chief component and progressively grab information from element and row/column relation tokens via a Separator Perceiver (Bézier-curve separator proposals), Components Correlator (decoupled cross-attention), and Structure Composer (sequence generation of separator indexes per cell), trained end-to-end with separator and structure losses and one-to-many separator assig
Key results
- On SciTSR-COMP-A, GrabTab reaches 94.0% F1, beating the second-best NCGM (89.5%) by about 4.5 points.
- On PubTabNet, GrabTab achieves 97.9 TEDS, above NCGM's 95.4 and other listed baselines.
- Ablation on SciTSR-COMP-A: adding Components Correlator raises F1 from 89.0 to 90.3, and the full Structure Composer version reaches 94.0 versus 91.6 for the rule-based variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GrabTab instead of TSRFormer | TSR on WTW · WTW | F1-score: 94.3 -> 95.1 | Helped |
| Use GrabTab instead of TSRFormer | TSR on SciTSR-COMP · SciTSR-COMP | F1-score: 98.9 -> 99.1 | Helped |
| Use GrabTab instead of NCGM | TSR on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 89.5 -> 94.0 | Helped |
| Use GrabTab instead of NCGM | logical structure recognition · TableBank | BLEU: 94.6 -> 95.0 | Helped |
| Use GrabTab instead of NCGM | logical structure recognition · PubTabNet | TEDS-Struct: 95.4 -> 97.9 | Helped |
| Add row/column relation bias (ablation) | Deform.-DETR on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 84.9 -> 89.0 | Helped |
| Add Components Correlator (ablation) | GrabT.w/o CC/MA/SC on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 89.0 -> 90.3 | Helped |
| Use one-to-many assignment instead of one-to-one assignment (ablation) | GrabTab without Structure Composer on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 90.3 -> 91.6 | Helped |
| Add Structure Composer (ablation) | GrabTab on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 91.6 -> 94.0 | Helped |
| Use curved separators instead of straight-line separators (ablation) | GrabTab on SciTSR-COMP-A · SciTSR-COMP-A | F1-score: 92.1 -> 94.0 | Helped |
Surprising findings
- Straight-line separator targets (GrabTab-S) perform worse than curved Bézier targets, especially on distorted tables.
- Implicit separators show higher attention divergence to table elements than explicit ones (0.29 vs 0.17), suggesting adjacent elements drive implicit separator recovery.
Limitations
- Candidate components (elements and relations) are extracted off-line rather than jointly trained.
- Computational complexity increases with the proposed deliberation modules.
- Relies on NCGM for candidate component extraction.
Experimental setup
Datasets: ICDAR-2013 ICDAR-2019 WTW UNLV SciTSR SciTSR-COMP SciTSR-COMP-A TableBank PubTabNet
Benchmarks: WTW SciTSR-COMP SciTSR-COMP-A TableBank PubTabNet ICDAR-2013 ICDAR-2019 UNLV SciTSR
Metrics: precision recall F1-score TEDS-Struct BLEU
Baselines: C-CTRNet FLAG-Net TSRFormer RobusTabNet TableFormer LGPMA NCGM
Builds on: Deformable DETR NCGM Pix2Seq Hungarian matching
Closest papers by meaning
- TDeLTA: A Light-Weight and Robust Table Detection Method Based on Learning Text Arrangement (AAAI 2024 · similarity 0.85)
- Decoupling Skeleton and Flesh: Efficient Multimodal Table Reasoning with Disentangled Alignment and Structure-aware Guidance (ICML 2026 · similarity 0.84)
- TabPedia: Towards Comprehensive Visual Table Understanding with Concept Synergy (NeurIPS 2024 · similarity 0.84)
- TABLET: A Large-Scale Dataset for Robust Visual Table Understanding (ICLR 2026 · similarity 0.84)
- TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding (ICLR 2026 · similarity 0.84)
- TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition (CVPR 2026 · 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.