TransMI: a Transfer-learning method for generalized Map Information evaluation

Pytorch implementation of our paper “TransMI: a Transfer-learning method for generalized Map Information evaluation”. We collect a Subjective dataset for deep learning in Information Theory of Cartography named SITC and propose a novel approach, dubbed TransMI, to measure the quality of generalized map information, which is faster, more unified, and more subjective.

1. Requirements

• pytorch
• torchvision
• numpy
• pandas

2. Dataset

We provide the Subjective dataset for deep learning in Information Theory of Cartography (SITC) in this repository.

File Name / Folder	Content
./data/pre_train/data	Relative SITC dataset
./data/pre_train/clsLabel.csv	Relative SITC labels
./data/fine_tune/data	Absolute SITC dataset
./data/fine_tune/MOS.csv	Absolute SITC MOS

2.1 Relative SITC (ReSITC)

ReSITC contains 2970 maps, among which there are 20 kinds of relative relationships as shown in the table below.

Label	Class
0	Increase in Polygon Type
1	Increase in Line Type
2	Increase in Point Type
3	Increase in Point Number
4	Increase in Polygon Number
5	Increase in Line Number
6	Distribution Tends to be Chaotic
7	Increase in Point Level
8	Increase in Polygon Complexity
9	Change in Color
10	Change in Point Meaning
11	Decrease in Polygon Type
12	Decrease in Line Type
13	Decrease in Point Type
14	Decrease in Point Number
15	Decrease in Polygon Number
16	Decrease in Line Number
17	Distribution Tends to be Orderly
18	Decrease in Point Level
19	Decrease in Polygon Complexity

2.2 Absolute SITC (AbSITC)

AbSITC contains 330 maps, and each map is scored by 15 participants according to the quality of generalized map information.

3. Reproducibility of the results

You can reproduce the results in the paper according to the following instructions.

3.1 Pre-training Stage

To reproduce the k-fold cross-validation results in the pre-training stage:

cd pre_train
python train.py

To generate the pretrained model for the fine-tuning stage:

cd ..
cd generate_ckpoint
python train.py

3.2 Fine-tuning Stage

To reproduce the k-fold cross-validation results in the fine-tuning stage:

cd ..
cd fine_tune
python train.py --model_path ../logs/generate_ckpoint_output/model.pth --model_id 1

4. Visualization

Here, we provide some prediction results of our model.