Research map / Detection and segmentation / Open-set and few-shot detection
Component-Level Segmentation for Oracle Bone Inscription Decipherment
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 a weakly supervised PVT-based encoder-decoder for component-level segmentation of oracle bone inscriptions, using few annotated and many weakly labeled characters plus expert stroke rules, evaluated with mIoU, auxiliary classifiers, and expert scoring.
Problem
Segmenting specific components within oracle bone inscription characters to help experts decipher undeciphered OBIs, where annotation is scarce, components resemble background strokes, and segmentation quality is hard to evaluate without expert knowledge.
Approach
Combines a small set of pixel-annotated component data with a larger set of category-only weakly labeled characters via feature-level anchor-positive-negative and anchor-positive-positive triplet losses, and adds expert stroke-rule classifiers (connected domain, circle, intersection counts) as constraints on the predicted segmentation.
Key results
- Ours achieves 61.45% mIoU and 75.50% recall, versus 58.76% mIoU for NAT and 53.01% for MCT.
- On expert recognition score, Ours reaches 3.31 (3.41 with noise removal) versus 1.81 for NAT and 1.48 for MCT.
- Ablation: removing the unannotated-data loss drops component accuracy from 33.91% to 33.04%, and removing stroke rules drops it to 24.03%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add unannotated data through Lunan (ablation) | Ours · OBI Component 20 test set | Component accuracy (%): 33.04 -> 33.91 | Helped |
| Add stroke-rule loss Lstroke (ablation) | Ours · OBI Component 20 test set | Component accuracy (%): 24.03 -> 33.91 | Helped |
| Remove strokes occupying less than 0.07% of image area (ablation) | Ours · OBI Component 20 test set | Recognition score: 3.31 -> 3.41 | Helped |
| Remove strokes occupying less than 0.07% of image area (ablation) | Ours · OBI Component 20 test set | Completeness score: 3.70 -> 3.59 | Hurt |
| Remove strokes occupying less than 0.07% of image area (ablation) | Ours · OBI Component 20 test set | Noise score: 2.55 -> 1.75 | Helped |
| Use Ours instead of Nat’23, Beit’23, or MCT’24 | OBI Component 20 test set · OBI Component 20 test set | mIoU (%): Nat’23 58.76; Beit’23 56.71; MCT’24 53.01; Ours 61.45 | Helped |
| Use Ours instead of Nat’23, Beit’23, or MCT’24 | OBI Component 20 test set · OBI Component 20 test set | recall (%): Nat’23 64.36; Beit’23 61.95; MCT’24 62.77; Ours 75.50 | Helped |
| Use Ours instead of Nat’23 or Beit’23 | OBI Component 20 test set · OBI Component 20 test set | precision (%): Nat’23 67.71; Beit’23 64.23; Ours 66.11 | Mixed |
| Use Ours instead of compared methods | OBI Component 20 test set · OBI Component 20 test set | Recognition score: Nat 1.81; Beit 1.81; MCT 1.48; Ours 3.31 | Helped |
Surprising findings
- Standard segmentation metrics (mIoU, precision, recall) are judged unsuitable for this task because thick strokes in the dataset cause correct segmentations to mismatch annotations.
- Auxiliary CDC, CC, and IC classifiers did not clearly discriminate between methods, though the component classifier aligned with visual and expert results.
Limitations
- Assumes each OBI contains only one target component, though many characters have multiple components.
- Standard metrics do not fit the thick-stroke annotation style, so evaluation relies on proxy classifiers and small-scale expert scoring.
- Auxiliary stroke-rule classifiers did not discriminate well between methods at test time.
Experimental setup
Datasets: OBI Component 20
Benchmarks: OBI Component 20 test split
Metrics: mIoU precision recall classification accuracy (component, CDC, CC, IC) expert scores (recognition, completeness, noise)
Baselines: BEiT NAT MCT
Builds on: Pyramid Vision Transformer triplet loss
Closest papers by meaning
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.90)
- Towards Automated Chinese Ancient Character Restoration: A Diffusion-Based Method with a New Dataset (AAAI 2024 · similarity 0.84)
- Making Visual Sense of Oracle Bones for You and Me (CVPR 2024 · similarity 0.83)
- ProtoSnap: Prototype Alignment For Cuneiform Signs (ICLR 2025 · similarity 0.82)
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.82)
- Advancing Manga Analysis: Comprehensive Segmentation Annotations for the Manga109 Dataset (CVPR 2025 · similarity 0.82)
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.