Research map / Detection and segmentation / CNN features
Document Structure Extraction using Prior based High Resolution Hierarchical Semantic Segmentation
Mausoom Sarkar, Milan Aggarwal, Arneh Jain, Hiresh Gupta, Balaji Krishnamurthy
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical high-resolution strip-wise segmentation (Highresnet) over low-res Lowresnet | Lowresnet · Forms Dataset: ChoiceGroup | MIoU (%): 72.61 -> 83.01 | Improved |
| add prior mask of previous strip when predicting next strip (ablation) | Highresnet · Forms Dataset: ChoiceGroup | MIoU (%): 80.14 -> 83.01 | Improved |
| add prior mask of previous strip when predicting next strip (ablation) | Highresnet · Forms Dataset: ChoiceGroup Title | MIoU (%): 78.89 -> 80.87 | Improved |
| use 1D dilated convs instead of 2D dilated convs (ablation) | Highresnet · Forms Dataset: ChoiceField | MIoU (%): 81.95 -> 84.05 | Improved |
| use 1D dilated convs instead of 2D dilated convs (ablation) | Highresnet · Forms Dataset: TextBlock | MIoU (%): 89.71 -> 90.55 | Improved |
| predict all four hierarchies jointly instead of one model per hierarchy (ablation) | Lowresnet · Forms Dataset: ChoiceGroup | MIoU (%): 70.55 -> 72.61 | Improved |
| predict all four hierarchies jointly instead of one model per hierarchy (ablation) | Lowresnet · Forms Dataset: ChoiceGroup Title | MIoU (%): 63.95 -> 69.03 | Improved |
| use a single shared final layer for all hierarchies instead of separate decoders (ablation) | Lowresnet · Forms Dataset: ChoiceGroup | MIoU (%): 69.37 -> 72.61 | Improved |
| use Highresnet instead of ImageNet-pretrained DeepLabV3+ | DLV3+ ImageNet · Forms Dataset: ChoiceGroup | MIoU (%): 56.11 -> 83.01 | Improved |
| use ImageNet pretraining for DeepLabV3+ (ablation) | DLV3+ · Forms Dataset: ChoiceGroup Title | MIoU (%): 18.69 -> 48.09 | Improved |
| use Lowresnet-TL instead of MSMT-FCN at stricter IoU 0.9 | MSMT-FCN · Marmot English (IoU 0.9) | table detection F1 (%): 45.9 -> 62.8 | Improved |
| use Lowresnet-TL instead of MSMT-FCN at IoU 0.8 | MSMT-FCN · Marmot English (IoU 0.8) | table detection F1 (%): 72.5 -> 73.7 | No clear effect |
Closest papers by meaning
- Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training (CVPR 2026 · similarity 0.84)
- DocSAM: Unified Document Image Segmentation via Query Decomposition and Heterogeneous Mixed Learning (CVPR 2025 · similarity 0.83)
- SRFUND: A Multi-Granularity Hierarchical Structure Reconstruction Benchmark in Form Understanding (NeurIPS 2024 · similarity 0.83)
- Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation (CVPR 2024 · similarity 0.83)
- Graph-based Document Structure Analysis (ICLR 2025 · similarity 0.83)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 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.
