Research map / Detection and segmentation / CNN features
Kernel Adaptive Convolution for Scene Text Detection via Distance Map Prediction
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 KAC, a scene text detector that predicts a text distance map with a plug-and-play Kernel Adaptive Convolution Module that mixes multiple kernel sizes with per-position adaptive weights, achieving strong accuracy and speed on curved/multilingual benchmarks.
Problem
Fixed-size convolution kernels limit feature extraction for scene text of variable scale and shape, and segmentation-based detectors that segment only text centers lose edge and direction information, while accurate two-stage or iterative methods sacrifice efficiency.
Approach
A Kernel Adaptive Convolution Layer runs parallel convolutions with kernel sizes 1,3,5,7 and weights them per spatial position using a predicted, normalized weight map; the detection head predicts a text distance map (plus segmentation) instead of a center region, with direction maps used only as training supervision.
Key results
- On Total-Text, KAC-ResNet50 reaches 90.98% F-measure at 25.0 FPS, versus TextBPN++ at 90.13% F and 13.2 FPS.
- On TD500, KAC-ResNet50 reaches 90.79% F-measure at 18.1 FPS, versus TextBPN++ at 90.10% F.
- On ICDAR19 ArT, KAC-ResNet50 reaches 80.0% F-measure at 4.6 FPS, below TextBPN++ (80.59% F), with the paper citing efficiency gains; on CTW1500 it reaches 86.84% F, below DPText-DETR (88.8%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add KACM to DB (ablation) | DB · TD500 | F-measure: 84.9 -> 86.1 | Helped |
| add KACM to KAC (ablation) | KAC · TD500 | F-measure: 89.76 -> 90.79 | Helped |
| replace text-region center segmentation with distance-map prediction (ablation) | KAC · Total-Text | F-measure: 88.42 -> 90.98 | Helped |
| add direction-map supervision (ablation) | Baseline · Total-Text | F-measure: 87.67 -> 88.09 | Helped |
| add segmentation-map supervision alongside direction-map supervision (ablation) | Baseline · Total-Text | F-measure: 88.09 -> 90.99 | Helped |
| vary KACM kernel count from n = 4 to n = 3 (ablation) | KACM · Total-Text | F-measure: 90.98 -> 90.54 | Hurt |
| vary KACM kernel count from n = 4 to n = 5 (ablation) | KACM · Total-Text | F-measure: 90.98 -> 90.78 | Hurt |
| vary KACM kernel count from n = 4 to n = 6 (ablation) | KACM · Total-Text | F-measure: 90.98 -> 90.82 | Hurt |
| use KAC-ResNet50 instead of TextBPN++ | scene text detection · Total-Text | F-measure: 90.13 -> 90.98 | Helped |
| use KAC-ResNet50 instead of TextBPN++ | scene text detection · TD500 | F-measure: 90.10 -> 90.79 | Helped |
| use KAC-ResNet50 instead of TextBPN++ | scene text detection · ICDAR19 ArT | F-measure: 80.59 -> 80.0 | Hurt |
| use KAC-ResNet50 instead of DPText-DETR | scene text detection · CTW1500 | F-measure: 88.8 -> 86.84 | Hurt |
Surprising findings
- On ICDAR19 ArT, the reported FPS of KAC (4.6) is lower than the efficiency claimed elsewhere, and the paper's headline 142.1% efficiency gain is relative to TextBPN++ (1.9 FPS) there.
- On CTW1500 the KAC-ResNet50 F-measure (86.84%) trails several recent methods, so the claimed advantage is efficiency rather than accuracy.
Limitations
- Accuracy on CTW1500 and ArT is competitive but not state-of-the-art.
- Relies on pretraining on SynthText and initialization from BPN++ backbone weights.
- Post-processing uses a binarization threshold tuned per dataset (0.3 to 0.475).
Experimental setup
Datasets: SynthText Total-Text MSRA-TD500 ICDAR19 ArT CTW1500
Benchmarks: Total-Text CTW1500 MSRA-TD500 ICDAR19 ArT
Metrics: precision recall F-measure FPS
Baselines: TextBPN++ DPText-DETR DB DB++ TextPMs TextBPN ContourNet PAN PSENet CT-Net
Builds on: DB BPN++ TextPMs KPN Dynamic Convolution
Closest papers by meaning
- A Text Detection System for Natural Scenes With Convolutional Feature Learning and Cascaded Classification (CVPR 2016 · similarity 0.89)
- ContourNet: Taking a Further Step Toward Accurate Arbitrary-Shaped Scene Text Detection (CVPR 2020 · similarity 0.88)
- Multi-Oriented Scene Text Detection via Corner Localization and Region Segmentation (CVPR 2018 · similarity 0.88)
- Arbitrary Shape Scene Text Detection With Adaptive Text Region Representation (CVPR 2019 · similarity 0.87)
- CentripetalText: An Efficient Text Instance Representation for Scene Text Detection (NeurIPS 2021 · similarity 0.87)
- LRANet: Towards Accurate and Efficient Scene Text Detection with Low-Rank Approximation Network (AAAI 2024 · similarity 0.85)
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.