Benchmark modern historical map vectorization processes

Abstract

Shape vectorization is a key stage of the digitization of high-scale historical maps, especially city maps. Having access to digitized buildings, building blocks, street networks and other typical from historical content opens many ways for historical studies: change tracking, morphological analysis, density estimations. In the context of the digitization of Paris atlases produced of the course of the 19th and early 20th centuries, we designed a processing pipeline capable of extracting closed shaped from historical maps, formalized as an instance segmentation problem, in an attempt to reduce the amount of manual work required. This pipeline relies on an edge filtering stage using deep filters, a closed shape extraction stage using a watershed transform, and a raster-to-vector conversion stage.

The following contributions are introduced:

• a public dataset over which an extensive benchmark is performed;
• a comparison of the performance of state-of-the-art deep edge detectors, among which vision transformers, and several deep and classical watershed approaches;
• a joint optimization of the edge detection and shape extraction stages;
• a study of the effects of augmentation techniques.

Results are made reproducible and reusable thanks to public data, code and results.

Usage

• Download

Download the dataset and put it into the folder Dataset

• Data loading

This repository contains:

📂IJGIS_benchmark_vector
 ┣ 📂benchmark         # Benchmark scripts for both datasets
 ┣ 📂config            # Config files for vision image transformers
 ┣ 📂data              # Dataset loader
 ┣ 📂dataset           # Datasets files including images and ground truths
 ┣ 📂demo              # Demos for the algorithms used in this paper
 ┣ 📂evaluation        # Evaluation code for pixel and topology evaluation
 ┃ ┣ 📂all_eval        # Evaluation code for historical map vectorization task
 ┃ ┃ ┣ 📂pixel_eval    # Pixel evaluation code
 ┃ ┃ ┣ 📂topo_eval     # Topology evaluation code 
 ┣ 📂inference         # Inferencing maps with trained models
 ┣ 📂licenses          # Licenses
 ┣ 📂loss              # Pixel and topology losses
 ┣ 📂model             # Pytorch models
 ┣ 📂pretrain_weight   # Pretrained weight for training
 ┣ 📂training          # Training scripts
 ┣ 📂utils             # Some utility files
 ┣ 📂watershed         # Watershed segmentation code {binary files + python version}

• Smart dataloader

In order to process maps with large size, we introduce Smart data loader which can divide the image into (overlap or non-overlap) patches, then feed it into pytorch dataloader.

• Training the maps

We seperated the training with topology losses and without topology losses into two different training files.

To start training:

cd training
python train_model_only.py --model_type <model_type> --cuda --gpu <gpu> --lr <learning_rate>

or

cd training
python train_loss.py --model_type <model_type> --cuda --gpu <gpu> --lr <learning_rate>

The training results is saved in the folder: training_info

📂training_info
 ┣ 📂 params                  # Saving every weights in the training
 ┣ 📂 reconstruction_png      # Reconstruct patch predictions into full maps
 ┣ <model>.txt                # Training logs

• Joint optimization validation maps

Joint optimization by grid search every predictions of epochs with Meyer waterhsed segmentation:

cd benchmark/
python run_bench_mws.py --image_dir ./training_info/<model_name>/reconstruct_png

or with binarization of EPM + edge filteirng (waterhsed segmentation with area=0, dynamics=0):

cd benchmark/
python run_bench_ccfilter.py --image_dir ./training_info/<model_name>/reconstruct_png

• Evaluation test maps

Evaluation the test maps and return pixel and topology evaluation results and save it into .json file.

cd inferencing/
python test_mws.py --cuda --gpu <gpu> --model_type <model type> -d <dyanmic value> -a <area value> --model <best model weight .pth file> 

• Infrencing unseen historical maps with watershed segmentation

Model_name and its related model_type (download weight file (in release) into the folder ./pretrain_weight/ ):

Model_name	model_type
hed_best_weight.pth	hed
hed_pretrain_best_weight.pth	hed_pretrain
bdcn_best_weight.pth	bdcn
bdcn_pretrain_best_weight.pth	bdcn_pretrain
mini_unet_best_weight.pth	mini-unet
mosin_best_weight.pth	mosin
topo_best_weight.pth	topo
bal_best_weight.pth	bal
pathloss_best_weight.pth	pathloss
vit_best_weight.pth	vit
pvt_best_weight.pth	pvt
unet_best_weight.pth	unet
unet_hws_best_weight.pth	unet
unet_aff_best_weight.pth	unet_aff
unet_bri_aff_best_weight.pth	unet_bri_aff
unet_bri_best_weight.pth	unet_bri
unet_bri_hom_best_weight.pth	unet_bri_hom
unet_tps_best_weight.pth	unet_tps
unet_bri_tps_best_weight.pth	unet_bri_tps
unet_hom_best_weight.pth	unet_hom
deep_watershed_best_weight.pth	dws

Map inferencing:

cd inferencing/
python new_map_inference.py --unseen --cuda --gpu <gpu> --model_type <model type> --model <best model weight .pth file>  --input_map_path <image path .jpg/.png file>

For example:

cd inferencing/
python new_map_inference.py --unseen --cuda --gpu 1 --model_type unet --model ./pretrain_weight/unet_best_weight.pth  --input_map_path ./BHdV_PL_ATL20Ardt_1898_0004-TEST-INPUT_color_border.jpg --vectorization