Skip to content

This library augments road images to introduce various real world scenarios that pose challenges for training neural networks of Autonomous vehicles. Automold is created to train CNNs in specific weather and road conditions.

License

Notifications You must be signed in to change notification settings

UjjwalSaxena/Automold--Road-Augmentation-Library

Repository files navigation

Automold

There are various types of image augmentations done to increase the image corpus for training neural networks. However for training CNNs to drive some special road conditions are required. These can be random gravels of the road or maybe snow. Rain and fog also reduce the visibility to a great extent. Automold helps in addressing these challenges and augments road images to have various weather and road conditions.

Importing road augmentation library Automold and helper functions library

import Automold as am
import Helpers as hp

Let's load up some images first

path='./test_augmentation/*.jpg'
images= hp.load_images(path)

visualize function helps in displaying images easily without requiring you to write the whole code.

hp.visualize(images, column=3, fig_size=(20,10))

png

Checking out the brighten function

brighten

parameters

image: image or image list

brightness_coeff amount of brightness (0<=brightness_coeff<=1), default: random

bright_images= am.brighten(images[0:3]) ## if brightness_coeff is undefined brightness is random in each image
hp.visualize(bright_images, column=3)

png

bright_images= am.brighten(images[0:3], brightness_coeff=0.7) ## brightness_coeff is between 0.0 and 1.0
hp.visualize(bright_images, column=3)

png

Let's darken a few images now

darken

parameters

image: image or image list

darkness_coeff amount of darkness (0<=darkness_coeff<=1), default: random

dark_images= am.darken(images[0:3]) ## if darkness_coeff is undefined darkness is random in each image
hp.visualize(dark_images, column=3)

png

dark_images= am.darken(images[0:3], darkness_coeff=0.7) ## darkness_coeff is between 0.0 and 1.0
hp.visualize(dark_images, column=3)

png

But what if you just want some random brightness or darkness in the images. Well try out the random_brightness function which receives an image or an image array

random_brightness

parameters

image: image or image list

dark_bright_images= am.random_brightness(images[4:7])
hp.visualize(dark_bright_images, column=3)

png

What about adding some shadows to the images.

add_shadow

Parameters

image: image or image list

no_of_shadows: no. of shadows, default: 1

rectangular_roi: (top-left x, top-left y, bottom-right x, bottom right y), default: lower half of image

shadow_dimension: no. of sides of the shadows (3<=shadow_dimension<=10), default: random

shadowy_images= am.add_shadow(images[4:7])
hp.visualize(shadowy_images, column=3)

png

shadowy_images= am.add_shadow(images[4:7], no_of_shadows=2, shadow_dimension=8)
hp.visualize(shadowy_images, column=3)

png

Now let's add some snow

add_snow

parameters

image: image or image list

snow_coeff: amount of snow (0<=snow_coeff<=1), default: random

snowy_images= am.add_snow(images[4:7]) ##randomly add snow
hp.visualize(snowy_images, column=3)

png

snowy_images= am.add_snow(images[4:7], snow_coeff=0.3)
hp.visualize(snowy_images, column=3)

png

snowy_images= am.add_snow(images[4:7], snow_coeff=0.8)
hp.visualize(snowy_images, column=3)

png

and now some rain

add_rain

parameters

image: image or image list

slant: deviation of rain from normal (-20<=slant<=20), default: random

drop_length: length of the drop (0<=drop_length<=100), default: 20 (pixels)

drop_width: width of the drop (1<=drop_width<=5), default: 1

drop_color: color of droplets, default: (200,200,200)

rain_type: values in 'drizzle','heavy','torrential', default: 'None'

rainy_images= am.add_rain(images[4:7])
hp.visualize(rainy_images, column=3)

png

rainy_images= am.add_rain(images[4:7], rain_type='heavy', slant=20)
hp.visualize(rainy_images, column=3)

png

rainy_images= am.add_rain(images[4:7], rain_type='torrential')
hp.visualize(rainy_images, column=3)

png

Note: drop_length and drop_width values are overriden when rain_type is not None

add_fog

parameters

image: image or image list

fog_coeff: amount of fog (0<=fog_coeff<=1), default: random

foggy_images= am.add_fog(images[4:7])
hp.visualize(foggy_images, column=3)

png

foggy_images= am.add_fog(images[4:7], fog_coeff=0.4)
hp.visualize(foggy_images, column=3)

png

foggy_images= am.add_fog(images[4:7], fog_coeff=0.9)
hp.visualize(foggy_images, column=3)

png

what about some gravels on the road now ?

add_gravel

parameters

image: image or image list

rectangular_roi: (top-left x, top-left y, bottom-right x, bottom right y), default: lower 3/4th of image

no_of_patches: no. of gravel patches required, default: 8

bad_road_images= am.add_gravel(images[4:7])
hp.visualize(bad_road_images, column=3)

png

bad_road_images= am.add_gravel(images[4:7], rectangular_roi=(700,550,1280,720),no_of_patches=20) ##too much gravels on right
hp.visualize(bad_road_images, column=3)

png

add_sun_flare

parameters

image: image or image list

flare_center: center coordinates (x,y) of the source, default: random

angle: angle of flare in radians, default: random

no_of_flare_circles: no. of secondary flare circles (0<=no_of_flare_circles<=20), default: 8

src_radius: radius of the primary flare source, default: 400 (pixels)

src_color: rgb color of the flare source and secondary circles, default: (255,255,255))

flare_images= am.add_sun_flare(images[4:7])
hp.visualize(flare_images, column=3)

png

import math
flare_images= am.add_sun_flare(images[4:7], flare_center=(100,100), angle=-math.pi/4) ## fixed src center
hp.visualize(flare_images, column=3)

png

add_speed

parameters

image: image or image list

speed_coeff: amount of speed (0<=speed_coeff<=1), default: random

speedy_images= am.add_speed(images[1:4]) ##random speed
hp.visualize(speedy_images, column=3)

png

speedy_images= am.add_speed(images[1:4], speed_coeff=0.9) ##random speed
hp.visualize(speedy_images, column=3)

png

add_autumn

parameters

image: image or image list

fall_images= am.add_autumn(images[0:3])
hp.visualize(fall_images, column=3)

png

fliph

parameters

image: image or image list

flipped_images= am.fliph(images[0:3])
hp.visualize(flipped_images, column=3)

png

flipv

parameters

image: image or image list

flipped_images= am.flipv(images[0:3])
hp.visualize(flipped_images, column=3)

png

random_flip

parameters

image: image or image list

flipped_images= am.random_flip(images[0:3])
hp.visualize(flipped_images, column=3)

png

add_manhole

parameters

image: image or image list

center: center of the ellipse (x,y), default: bottom center of the image

color: rgb tuple, default: if type parameter not defined: (67,70,75), else: default color mentioned in type.

height: vertical dimension of the hole, int , default: 25th portion of the image height.

width: horizontal dimension of the hole, int, default: 3/25th portion of the image height.

type: values in 'closed','open', default: 'closed'

manhole_images= am.add_manhole(images[0:3])
hp.visualize(manhole_images, column=3)

png

correct_exposure

parameters

image: image or image list

exposure_images= am.correct_exposure(images[0:3])
hp.visualize(exposure_images, column=3)

If a series of augmentations is required from above types augment_random function can be handy

augment_random

image: image or image list

aug_types: list of Automold functions, eg: ['add_snow','add_rain'], default: all aug functions are executed

volume: 'same' or 'expand', default: expand

        same: keeps the volume of corpus unchanged, applies random aug_types on images

        expand: applies all aug_types on all images and expands output corpus
aug_images= am.augment_random(images[4:6], volume='same')  ##2 random augmentations applied on both images
hp.visualize(aug_images,column=3,fig_size=(20,20))

png

aug_images= am.augment_random(images[4:6], volume='expand')  ##all aug_types are applied in both images
hp.visualize(aug_images,column=3,fig_size=(20,20))

png

aug_images= am.augment_random(images[4:6], aug_types=['add_sun_flare','add_speed','add_autumn'], volume='expand')  ##all aug_types are applied in both images
hp.visualize(aug_images,column=3,fig_size=(20,10))

png

aug_images= am.augment_random(images[4:6], aug_types=['add_sun_flare','add_speed','add_autumn'], volume='same')  ##2 random aug_types are applied in both images
hp.visualize(aug_images,column=3,fig_size=(20,10))

png

Performance statistics

aug_types=["random_brightness","add_shadow","add_snow","add_rain","add_fog","add_gravel","add_sun_flare","add_speed","add_autumn","random_flip","add_manhole"]
dict_time={}
import time
for aug_type in aug_types:
    t=time.time()
    command='am.'+aug_type+'(images)'
    result=eval(command)
    dict_time[aug_type]=time.time()-t
    t=time.time()
print('Average Time taken per augmentaion function to process 1 image:')
tot=0
for key, value in dict_time.items():
    tot+=value
    print(key, '{0:.2f}s'.format(value/len(images)))

print('-----------------------')
print('Total no. of augmented images created:', len(aug_types)*len(images))
print('-----------------------')
print('Total time taken to create ',len(aug_types)*len(images),' augmented images:', '{0:.2f}s'.format(tot))
        Average Time taken per augmentaion function to process 1 image:
        add_rain 0.02s
        add_sun_flare 0.09s
        add_fog 0.37s
        add_speed 0.20s
        random_brightness 0.05s
        add_shadow 0.01s
        random_flip 0.00s
        add_manhole 0.01s
        add_autumn 0.31s
        add_gravel 0.04s
        add_snow 0.06s
        -----------------------
        Total no. of augmented images created: 99
        -----------------------
        Total time taken to create  99  augmented images: 10.42s

Note: The load_images helper function resizes all images to 1280x720 and thus all returned images have the same dimensions. If original image dimensions are required please write a new load function

Some more helpful functions are in pipeline and will get added to the library asap. Thanx


MIT License

Copyright (c) 2018 Ujjwal Saxena

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

About

This library augments road images to introduce various real world scenarios that pose challenges for training neural networks of Autonomous vehicles. Automold is created to train CNNs in specific weather and road conditions.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published