svilav/polygon_generator

Generate random simple polygons (with augmentation)

Polygon generator

• The algorithm starts with the random point, then creates the polygon by sampling points on a circles centered at the point.
• Randon variance is added by random angles between sequential points and random radii from the centre to each point.
• Additional verifications are performed to ensure simple polygons and no collinearity between adjacent vertex. Points are in clockwise order. In most cases, first point will be somwhere in the mid-left.
• Images are generated with random nubmer of the following distortions:

1. GaussianBlur(sigma=(0.0, 4.0)),

2. MotionBlur(k=15),

3. MultiplyElementwise((0.5, 1.5), per_channel=0.5), # Noise inside figure

4. Multiply((0.5, 1.5)), #Brightness

5. Dropout(p=(0, 0.2)), #Random pxls to 0

6. CoarseDropout((0.0, 0.05), size_percent=(0.02, 0.25)), #Big rectangular dropout

7. BlendAlphaSimplexNoise(iaa.EdgeDetect(1.0), sigmoid_thresh=iap.Normal(10.0, 5.0)), #Blurred blobs

8. BlendAlphaHorizontalLinearGradient(iaa.AveragePooling(11), start_at=(0.0, 1.0), end_at=(0.0, 1.0)), #Massive blend

9. PiecewiseAffine(scale=(0.01, 0.02)), #Tiny curve

10. AveragePooling([2, 8])

• Under defualt setting, generation of 1k images takes 14 sec in Colab.

Config:

Parameter	Description
img_num:	how many images to generate
distort:	whether to apply image distortion
numVerts:	number of polygon vertex
dir:	dir containing 'images" folder
collin_coef:	[0,1] how far angles shall be from 180°
aveRadius:	mean for sampling radii
angular_variance:	[0,1] the variance in angles
radii_variance:	[0,1] the variance in length
input_size:	image size will be (input_size,input_size)

Installation:

Clone this repo

git clone https://username:[email protected]/Sviatoslav-Lavrinchuk/polygon_generator.git

cd polygon_generator

PIP

pip install --upgrade -r requirements.txt

Run

First edit "poly_gen_config.yaml".

Then run: python poly_gen.py

Project structure

poly_gen.py

poly_gen_config.yaml

dir

images