Research map / Graph learning / Knowledge and temporal graphs
SpaceGTN: A Time-Agnostic Graph Transformer Network for Handwritten Diagram Recognition and Segmentation
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 SpaceGTN, a time-agnostic graph transformer that recognizes and segments online handwritten diagrams from spatial stroke information only, and releases the OHSD free-drawing dataset.
Problem
Online handwritten diagram recognition and symbol segmentation degrade when stroke temporal information is missing, modified, or unreliable (e.g., restored offline strokes), and existing datasets use templated drawings.
Approach
A graph transformer (GTN-based) over stroke graphs that uses only (x,y) stroke coordinates, with a dynamic graph built from spatial proximity, K-dhop connections, and image/shape/curvature similarity scores, plus edge and spatial encodings added to attention.
Key results
- SpaceGTN reaches 98.13% SCA on CASIA-OHFC and 99.78% SCA on OHSD without temporal information, beating Inst-GNN (92.86% and 95.52%).
- On FC_A, SpaceGTN achieves 99.02% SCA versus 98.44% for Inst-GNN.
- Ablation on CASIA-OHFC: full model with Kdhop, dynamic graph building, edge and spatial encodings reaches 98.13% SCA, up from 94.06% for the baseline configuration.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace InstGNN with SpaceGTN | online handwritten diagram recognition · CASIA-OHFC | SCA: 92.86 -> 98.13 | Helped |
| replace InstGNN with SpaceGTN | online handwritten diagram recognition · OHSD | SCA: 95.52 -> 99.78 | Helped |
| add Kdhop graph building without dynamic graph building (ablation) | SpaceGTN · CASIA-OHFC | SCA: 94.06 -> 94.56 | Helped |
| add dynamic graph building without Kdhop graph building (ablation) | SpaceGTN · CASIA-OHFC | SCA: 94.06 -> 96.13 | Helped |
| add edge encoding without graph-building or spatial-encoding modules (ablation) | SpaceGTN · CASIA-OHFC | SCA: 94.06 -> 95.90 | Helped |
| add spatial encoding without graph-building or edge-encoding modules (ablation) | SpaceGTN · CASIA-OHFC | SCA: 94.06 -> 95.02 | Helped |
| use Kdhop and dynamic graph building together (ablation) | SpaceGTN · CASIA-OHFC | SCA: 94.06 -> 97.42 | Helped |
| add edge encoding to Kdhop and dynamic graph building (ablation) | SpaceGTN · CASIA-OHFC | SCA: 97.42 -> 97.81 | Helped |
| add spatial encoding to Kdhop and dynamic graph building (ablation) | SpaceGTN · CASIA-OHFC | SCA: 97.42 -> 97.76 | Helped |
| add both edge encoding and spatial encoding to Kdhop and dynamic graph building (ablation) | SpaceGTN · CASIA-OHFC | SCA: 97.42 -> 98.13 | Helped |
Surprising findings
- Competing online methods such as Inst-GNN and EGAT drop markedly when temporal information is removed, confirming their dependence on timing.
- About 70% of OHSD diagrams and 40% of symbols contain stroke modifications, with free drawing producing 20-30% more modified strokes than template drawing.
Limitations
- Ablation and main comparisons are reported on CASIA-OHFC and OHSD, with the strongest gains on OHSD; generalization beyond these diagram types is not shown.
- Baselines lacking public code had to be reimplemented by the authors, which may affect fairness of comparisons.
Experimental setup
Datasets: FA FC_A FC_B CASIA-OHFC OHSD
Benchmarks: FC_A FC_B CASIA-OHFC OHSD
Metrics: Stroke Classification Accuracy Stroke Classification Precision Symbol Recognition F1-Score
Baselines: DETR Deformable DETR Mask R-CNN Faster R-CNN ORSAD EGAT Inst-GNN
Builds on: GTN (Ying et al. 2021) PNAConv Inst-GNN Shape Context
Closest papers by meaning
- SSAN: A Symbol Spatial-Aware Network for Handwritten Mathematical Expression Recognition (AAAI 2025 · similarity 0.82)
- Enhance Sketch Recognition’s Explainability via Semantic Component-Level Parsing (AAAI 2024 · similarity 0.81)
- SketchINR: A First Look into Sketches as Implicit Neural Representations (CVPR 2024 · similarity 0.81)
- Efficient Unlearning for Spatio-temporal Graph (Student Abstract) (AAAI 2025 · similarity 0.81)
- SignGraph: A Sign Sequence is Worth Graphs of Nodes (CVPR 2024 · similarity 0.80)
- Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch (CVPR 2025 · similarity 0.80)
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.