Capturing More: Learning Multi-Domain Representations for Robust Online Handwriting Verification
ACM International Conference on Multimedia (ACM MM), 2025, Oral
⭐Official code of the SPECTRUM model.
SPECTRUM is an online handwriting verification model, designed to integrate temporal and frequency features from a micro-to-macro level to enrich personal handwriting representations.
git clone https://github.com/NiceRingNode/SPECTRUM.git
cd SPECTRUM
conda create -n spectrum python=3.8.16
conda activate spectrum
pip install -r requirements.txt
Download the MSDS-ChS, MSDS-TDS, and DeepSignDB datasets, and prepare the .pkl files for training and testing.
The preprocessed data should be placed at the data folder.
Run the following code to conduct training on the MSDS-ChS dataset:
python train.py --data_name signature --name msdschs --gpu 0
Run the following code to conduct training on the MSDS-TDS dataset:
python train.py --data_name real --name msdstds --gpu 0
Run the following code to conduct training on the DeepSignDB dataset:
python train.py --data_name deepsigndb --name deepsign --gpu 0
One can specify the running devices using the --gpu parameter.
The checkpoints should be saved in the weights folder.
For testing on the MSDS-ChS and MSDS-TDS datasets, using the following command and replace folder to the folder name of the tested checkpoint (e.g., weights/20251212-171546-msdschs).
python test.py --weights weights/{folder} --epoch 39
For testing on the DeepSignDB dataset, first change to the deepsign directory.
cd deepsign
Then, specify the checkpoint's folder (e.g., 20251212-181546-deepsign, no need to type the weights) using the --weights parameter in the eval.sh file and run this file for evaluation.
bash eval.sh
The results reported in the paper correspond to those after "Overall EER under global threshold".
@inproceedings{spectrum2025zhang,
author = {Zhang, Peirong and Ding, Kai and Jin, Lianwen},
title = {{Capturing More: Learning Multi-Domain Representations for Robust Online Handwriting Verification}},
year = {2025},
booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM)},
pages = {1471–1479},
numpages = {9},
}Peirong Zhang: eeprzhang@mail.scut.edu.cn
Copyright 2025, Deep Learning and Vision Computing (DLVC) Lab, South China China University of Technology. http://www.dlvc-lab.net.