Dataset annotation plateform for deep learning segmentation

Python interface to manually segment a dataset. In this version, images must be in .jpg format and the target array in the .npy format.

Download the requirements

foo@bar:~$ pip install -r requirements.txt

Config file description

Build your config.csv file to parameter your different classes

• class column : name of the class you want to define
• bgr_color column : the color you wnat to display your class in BGR format
• shape column : the shape of drawing
  • 0 : draw a polygone area by placing countour points (Then press enter to validate)
  • 1 : pixel by pixel drawing
  • 2 : draw filled circles by selecting two point on it
  • 3 or plus : draw a polygone with defined number of countour points (Automatically validate at the number of points)

My config.csv file to segment a pool table environment is the following

class	bgr_color	shape
Cue Stick	#004080	0
Red Ball	#0000FF	2
Yellow Ball	#00FFFF	2
Black Ball	#999999	2

Run the interface

foo@bar:~$ python main.py train_sample target_sample config.csv

COMMANDS

• Left button : reference a point on the current class channel or erase if the brush eraser is activated
• Enter key : validate the current references points in the case of type >= 3
• U key : undo the last reference point
• B key: activate/deactivate the brush eraser
• E key : go to the next class channels
• D key : go to the previous class channels
• F key : go to the next image/target
• S key : go to the previous image/target
• Z key : navigate between zoom modes (No zoom, x2, x3, x4, x5)
• Return key : delete the current sample
• Q key : exit the program

As you can see I choose my keys according to my AZERTY keyboard and my personnal wishes. You can change by yourself the shortcuts. At the top of the main.py file I set a dictionnary of keys to run the program.

class objectview(object):
    def __init__(self, d):
        self.__dict__ = d


keys = {
    'next_channel': ord('e'),           # ASCII number = 101
    'previous_channel': ord('d'),       # ASCII number = 100
    'next_image': ord('f'),             # ASCII number = 102
    'previous_image': ord('s'),         # ASCII number = 115
    'zoom': ord('z'),                   # ASCII number = 122
    'validate': 13,                     # ASCII number = 13
    'undo': ord('u'),                   # ASCII number = 117
    'brush': ord('b'),                  # ASCII number = 98
    'delete': 8,                        # ASCII number = 8
    'quit': ord('q')                    # ASCII number = 113
}

# to transform the dictionary into an object
keys = objectview(keys)

You can changes theses values according to the ASCII table. You can see that I had created an objectview() class. It permits to have a more lisible script, transforming the dictionnary into an object and attributes. It replaces the access from keys['next_image'] to keys.next_image for example.