WatermarkReco

Pytorch implementation of Paper "Large-Scale Historical Watermark Recognition: dataset and a new consistency-based approach"

[Arxiv] [Project website]

The project is an extension work to ArtMiner. If our project is helpful for your research, please consider citing :

@inproceedings{shen2019watermark,
          title={Large-Scale Historical Watermark Recognition: dataset and a new consistency-based approach},
          author={Shen, Xi and Pastrolin, Ilaria and Bounou, Oumayma and Gidaris, Spyros and Smith, Marc and Poncet, Olivier and Aubry, Mathieu},
          booktitle={Arxiv},
          year={2019}
        }

Table of Content

• Installation
• Classification
• Local Matching
• Feature Learning
• Visual Results

Installation

Dependencies

Code is tested under Pytorch > 1.0 + Python 3.6 environment. To install all dependencies :

bash requirement.sh

Datasets

We release our watermark dataset composed of 4 subsets targeting on 4 different tasks: classification, one-shot, one-shot cross-domain and large-scale one-shot cross-domain recognition.

You can run the following command to directly download the dataset:

cd data/
bash download_data.sh ## Watermark + Shoes / Chairs datasets

Or click here(~400M) to download it.

A full description of dataset is provided in our project website.

Models

To download pretrained models:

cd model/
bash download_model.sh # classification models + fine-tuned models

Classification

Dataset: A Train

cd classification/
bash demo_train.sh # Training with Dropout Ratio 0.7

Local Matching

One-shot Recognition

Dataset: A Test

cd localMatching/
bash demo_A.sh 

Feature Similarity Baselines:

cd featComparisonBaseline/
bash bestParam.sh # Run with resolution 256 * 256
bash run.sh # Run with different resolutions

One-shot Cross-domain Recognition

Dataset: B Test

cd localMatching/
bash demo_B.sh # Using drawing or synthetic as references with / without finetuned model

Dataset: Shoes / Chairs

cd localMatching/
bash demo_SBIR.sh # Evaluate on Shoes and Chairs dataset with / without finetuned model

Large Scale One-shot Cross-domain Recognition (16,753-class)

Dataset: B Test + Briquet

cd localMatching/
bash demo_Briquet_Baseline.sh # AvgPool, Concat and Local Sim. baselines
demo_Briquet_Ours.sh # Our approach w / wo F.T.

Feature Learning

Dataset: B Train

cd featureLearning/
bash demo_B_Finetune.sh # Eta = 3 for both drawing and synthetic references

Dataset: Shoes / Chairs

cd featureLearning/
bash demo_SBIR_Finetune.sh # Eta = 4 for chairs and shoes

Visual Results

More visual results can be found in our project website.

Top-5 retrieval results on Briquet + B Test dataset with using engraving as references:

Top-5 retrieval results on Briquet + B Test dataset with using synthetic image as references:

Top-5 retrieval results on Shoes / Chairs Test dataset:

Acknowledgment

This work was partly supported by ANR project EnHeritANR-17-CE23-0008 PSL Filigrane pour tous project and gifts from Adobe to Ecole des Ponts.