---

tags: SoDUCo

---

<style>
.reveal {
font-size: 24px;
}
img[alt=im] {
width: 500px;
}
img[title~="reduced"] { width: 80%;}
</style>

---

# Historical map vectorization

---
## ACTIONS
- create gitlab repo
- add extra attribute filter in HW
- setup pipeline for deep watershed experiment
- could it be possible to have some tests on BSDS500?
- interactive tool?

---

## Choice of Journal

This table is our proposition of journal in which publish, ordered by preference.

| Journal | Impact Factor | Rank | Review duration |
| ------- | ------------- | ----- | ------------ |
| PR | 7.74 | Q1 | fast |
| CVIU | 3.876 | Q1 | very slow |
| **PRL** | **3.756** | **Q1/Q2** | **< 20 weeks** |
| JMIV | 1.627 | Q2/Q3 | (out of scope) |
| IJDAR | 2.085 | Q3 | slow |

PRL option:
- https://www.editorialmanager.com/prletters/default1.aspx
- https://www.elsevier.com/journals/pattern-recognition-letters/0167-8655/guide-for-authors
- Gold Open Access fees: USD 2,270
- 7 pages double columns (pretty concise)

---

## The research question?

- How to correct filtering the unnecessary information from the historical maps? (Deep learning in semantic edge extraction)
- How to gurantee closed shapes from the filtering edge images (EPMs)? (Mayer and hierarchy watershed)
- How to do interactive segmentation to involve human in correcting vectors? (Interactive segmentation in hierarchy watershed)
- What happen if we have invalid jordan polygon?

---

## Pipeline of the Journal

### DGMM 2021 extended pipeline

Vectorization is added


### Our proposal

Using hierarchical watershed


### Produced vectorized lines

Illustration

---

## Planning of the paper

### 1, Finding the best strategies for picking the best EPMs

- Choosing the best weights according to the COCO panoptic when connected component labelling @ 0.5 - BASELINE
- Test different variant of the deep edge extraction methods (U-Net, HED, BDCN, ConnNet)

---

### 2, Test different attributes for best segmentation quality

- Area Attributes (range from 100: 1000)
- Dynamic attributes (range from 0: 10)
- Area + Dynamic attributes (A combination of both) [^1]
- Other attributes, more efficient for this task

[^1]: https://hal.archives-ouvertes.fr/hal-01552420/document

---

### 3, Map vectorization through watershed lines and saliency maps

The vectorization algotirhm uses inputs:
- Watershed lines
- Saliency maps
to vectorized historical maps.

---

### 4, Compare the watershed results with deep watershed

Instead of using watershed on EPM, the extra experiment on using deep watershed methods to predict watershed levels should be compare to Mayer+hierarchy watershed results.

---

### 5, A solid evaluation protocols evaluate both segmenation quality and vectorization quality in pixel-level, topological-level.

- Pixel: Dice, Soft precision, recall and clDIce
- Topology: Betti, COCO panoptic, ARI, VOI

---

### 6, Statistical analysis on objects

- The statistic in size in objects
- The statistical relation between the size and the correctness of the objects

---

### 7, The interactive segmentation toolkit for correcting historical map vectors

- The reason why we use the hierarchical watershed

---

### 8, Non horizontal cut for optimal segmentation

- Another reason to use the hierarchical watershed