Graffiti Dataset Toolbox

This is toolbox I use to work with my graffiti dataset.

Graffiti dataset

In this repository are few samples of this dataset to better explain its functionality.

Source Image	Graffiti Mask	Background mask	Background graffiti	Incomplete Graffiti	Tool used	Number of colors
					Spray paint	Many
					Spray paint	One
					Spray paint	One
					Spray paint	One
					Spray paint	One

All samples have following attributes

• Readable text - If graffiti contains some readable text
• Tool used - Tool used to draw the piece, currenty this is can be marker or spray_paint
• Number of colors used - 0 if more than one color is used, 1 othervise
• GPS coordinates - GPS coordinates extracted from EXIF data
• Graffiti Type - tag or graffiti

All samples have following masks

• Graffiti Mask - Mask defining main graffiti in the image
• Background mask - Mask defining background. This shouldn't contain any graffiti pixels
• Background graffiti - Samples sometimes contain graffiti other than the main one. This mask define those
• Incomplete Graffiti - This can define parts with missing or damaged areas of graffiti.

Data augmentation

This toolbox implements few augmentations which can be applied to dataset samples

Elastic transformation

Source Image	Transformed image

Modified colors

Source Image	Modified image

Modified colors + Elastic transformation

Source Image	Transformed image

Random background

Source Image	Graffiti on random background

Random background + Random color

Source Image	Transformed image

All transformations

Source Image	Transformed image

Features

Samples have multiple features attached to it

Graffiti colors

The most common colors in the graffiti

Source Image	RGB DBSCAN clusters

from graffiti_dataset.dataset import DatasetSample
from graffiti_dataset.tools import draw_main_colors

sample = DatasetSample('./dataset/graffiti_sample/8703f3c389a1f73f.p')

sample_main_colors = sample.main_colors()

draw_main_colors(sample_main_colors, 'colors.png')

GPS coordinates

GPS coordinates obtained from EXIF

Locations of samples in this repository :

How to generate such map

dataset_samples = glob.glob('./dataset/graffiti_sample/*.p')

gps_coordinates = []

for sample_path in dataset_samples:
    sample = DatasetSample(sample_path)
    gps_coordinates.append([sample.gps_latitude, sample.gps_longitude, sample.sample_id])

draw_map(np.array(gps_coordinates), f'./readme_images/map.html')

Color analysis

3D cube representing colors in sample can be made with tools.draw_color_cube(sample, 'pixels.html')

Colors represented as HSV can be desplayed with tools.draw_hsv_pixels(sample, 'out.png')

Process data from LabelBox

python3 labelbox_to_pickle.py export.csv dataset/graffiti dataset.csv

• export.csv - CSV file exported from LabelBox
• dataset/graffiti - Directory where processed samples should be stored
• dataset.csv - CSV file made when files were preprocessed before uploading to LabelBox. We need to do this in order to save GPS coordinates

Prepare data to upload to LabelBox

python3 data_to_labelbox.py './inout_data/*.JPG' ./output_direcotry/ ./new_data.csv

• ./inout_data/*.JPG - Where to look for new data
• ./output_direcotry/ - Where to put processed images
• ./new_data.csv - Where to store CSV describing new data