stillleben.camera_model module

Introduces camera noise effects.

Reference
- Functions

This module systematically models the different noise sources present in modern digital cameras.

The implementation largely follows the work of A. Carlson et al.: Modeling Camera Effects to Improve Visual Learning from Synthetic Data (2018), arXiv

Typical usage

import stillleben as sl
import torch

# Input image (CHW)
img = torch.rand(3, 1080, 1920)

# Everything is much faster on GPU!
img = img.cuda()

# Apply noise
img_with_noise = sl.camera_model.process_image(img)

Functions

def chromatic_aberration(rgb, translations, scaling): Introduces chromatic aberration effects.
def blur(rgb, sigma): Applies Gaussian blur
def exposure(rgb, deltaS): Re-exposes the image
def noise(rgb, a, b): Applies additional Poissonian-Gaussian noise
def color_jitter(tensor_img, hue_shift): Hue shift.
def process_deterministic(rgb, chromatic_translation, chromatic_scaling, blur_sigma, exposure_deltaS, do_noise, noise_a, noise_b, hue_shift): Process image with given noise model parameters.
def process_image(rgb): Process image with random noise parameters.

Function documentation

def stillleben.camera_model.chromatic_aberration(rgb, translations, scaling)

Introduces chromatic aberration effects.

Parameters
rgb	3xHxW input RGB image
translations	3x2 translation tensor (tx, ty) for each of R, G, B
scaling	[sr, sg, sb] scaling factor for each of R, G, B
Returns	3xHxW output tensor

def stillleben.camera_model.blur(rgb, sigma)

Applies Gaussian blur

Parameters
rgb	3xHxW input RGB image
sigma	Standard deviation of the Gaussian
Returns	3xHxW output RGB image

def stillleben.camera_model.exposure(rgb, deltaS)

Re-exposes the image

Parameters
rgb	3xHxW input RGB image, float [0-1]
deltaS	exposure shift (usually -1 to 1)
Returns	3xHxW output image

def stillleben.camera_model.noise(rgb, a, b)

Applies additional Poissonian-Gaussian noise

Parameters
rgb	3xHxW input RGB image, float [0-1]
a	variance factor of the signal-dependant noise (var = a*y(x))
b	standard deviation of the signal-independant noise
Returns	modified image

See Foi, Alessandro, et al. (2008): Practical Poissonian-Gaussian noise modeling and fitting for single-image raw-data, PDF

def stillleben.camera_model.color_jitter(tensor_img, hue_shift)

Hue shift.

Parameters
tensor_img	3xHxW RGB float tensor [0,1]
hue_shift	hue shift to apply (-0.5 to 0.5)
Returns	jittered image