segmentation-evaluation

Segmentation evaluation tool


Keywords
segmentation-evaluation, segmentation, evaluation, image, analysis, image-analysis, ground-truth
License
GPL-3.0
Install
pip install segmentation-evaluation==1.0.4

Documentation

Segmentation Evaluation Tool

This tool has two main characteristics:

  1. manually segment an image in order to create the ground truth
  2. evaluate a segmentation through a ground truth in terms of recall and precision

The documentation is reachable through that link.

GitHub release (latest by date) PyPI PyPI - Python Version GitHub GitHub All Releases GitHub stars GitHub forks

Table of Content

Installation

Use pip to install the tool:

pip install segmentation-evaluation

Usage

In order to use the tool, open the terminal and digit segmentation-evaluation:

$ segmentation-evaluation 

After that a printed menu is shown:


PERFORMANCE TOOL

  1. Press 1 to manually segment the image
  2. Press 2 to evaluate programmatically segmentation

In order to exit from the menu press ctrl + c.

Segmentation

Choosing 1 from the menu, the segmentation mode is activated.
After inserted the absolute path of the image and the *absolute path of the folder in which save the results, a window that shows the required image is displayed.

In order to insert a point DOUBLE CLICK in the desired position.
A green point is displayed. Then, inserting another point the segment between the two points is displayed.
After the third insertion of a point a polygon in which vertexes are the points is shown.

A point A line
A polygon with 3 vertexes with 4 vertexes

In addition to inserting new points it is possible to perform several action inserting a specific char while the image window is shown.
The possible actions are:

Char Action
u Undo last inserted point
c Clear all inserted points (not removes the already saved polygons)
s Save points and go to execution to insert label
q Terminate the insertion

Once the saving action is a pop up message asks to return to the terminal window where it is possible to insert the label associated to the polygon.
Then a printed string asks to come back to the image window to insert new polygons.

If you don't want to insert new polygon, come back to the image window and type q. A window shows all objects inserted associated to the label and typing again q the execution will terminate.

Evaluation

In order to perform the evaluation 3 paths need provided:

  1. path to the JSON file representing the correct segmentation, the ground truth
  2. path to the JSON file representing the segmentation to evaluate
  3. path to the folder in which save the evaluation

The evaluation is performed computing the recall and precision values. In order to consider a match good, the object has to satisfy an intersection over the union greater or equal than 80%.

The results are saved into a json file containing several values, the associated key is the label matched in both json files as input. In addition a final value that collects overall parameters associated to the key overall_performance.
Each item of the dictionary is another dictionary containing these keys:

  • "n_good_match": number of polygons matched between the two json files
  • "n_correct": number of polygons present in the ground truth
  • "n_found": number of polygons present in the segmentation to evaluate
  • "recall": n_good_match / n_correct
  • "precision": n_good_match / n_found

NB: the structure of the inputted json must be compliant with the following structure:

{
    "0": {
        "label": "name",
        "points": [
            [803, 1138],
            [980, 1167],
            [972, 1370],
            [797, 1331]
        ]
    },
    "1": {
        "label": "",
        "points": []
    }
}

The main keys is an incremental value that identify uniquely the polygon.
Each polygon must contain a key "label" with the name (as string) and "points" that has a list as value, the list contains the vertexes of the polygon. Each vertex is a point represented by a list or a tuple of the (x, y) coordinate.

Version

  • 1.0: first release

Credits

The tool was written by:

  • Mattia Sanchioni
  • Alessandro Concetti

License

This project is under the GPL-3.0 license - see the LICENSE.md file for more details