experiments/overlap/augmentations/distortion.py [24:96]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - softening = 1 return { 'time' : time, 'size' : size, 'eta' : eta, 'lens_scale' : lens_scale, 'lighting_amount': lighting_amount, 'softening' : softening} def transform(self, image, time, size, eta, lens_scale, lighting_amount, softening): def caustic_noise_kernel(point, time, size): point = point / size p = (point % 1) * 6.28318530718 - 250 i = p.copy() c = 1.0 inten = 0.005 for n in range(5): t = time * (1.0 - (3.5 / (n+1))) i = p + np.array([np.cos(t-i[0])+np.sin(t+i[1]),np.sin(t-i[1])+np.cos(t+i[0])]) length = np.sqrt((p[0] / (np.sin(i[0]+t)/inten))**2 + (p[1] / (np.cos(i[1]+t)/inten))**2) c += 1.0/length c /= 5.0 c = 1.17 - c ** 1.4 color = np.clip(np.abs(c) ** 8.0, 0, 1) return np.array([color, color, color]) def refract(incident, normal, eta): if np.abs(np.dot(incident, normal)) >= 1.0 - 1e-3: return incident angle = np.arccos(np.dot(incident, normal)) out_angle = np.arcsin(np.sin(angle) / eta) out_unrotated = np.array([np.cos(out_angle), np.sin(out_angle), 0.0]) spectator_dim = np.cross(incident, normal) spectator_dim /= np.linalg.norm(spectator_dim) orthogonal_dim = np.cross(normal, spectator_dim) rotation_matrix = np.stack((normal, orthogonal_dim, spectator_dim), axis=0) return np.matmul(np.linalg.inv(rotation_matrix), out_unrotated) def luma_at_offset(image, origin, offset): pixel_value = image[origin[0]+offset[0], origin[1]+offset[1], :]\ if origin[0]+offset[0] >= 0 and origin[0]+offset[0] < image.shape[0]\ and origin[1]+offset[1] >= 0 and origin[1]+offset[1] < image.shape[1]\ else np.array([0.0,0.0,0]) return np.dot(pixel_value, np.array([0.2126, 0.7152, 0.0722])) def luma_based_refract(point, image, caustics, eta, lens_scale, lighting_amount): north_luma = luma_at_offset(caustics, point, np.array([0,-1])) south_luma = luma_at_offset(caustics, point, np.array([0, 1])) west_luma = luma_at_offset(caustics, point, np.array([-1, 0])) east_luma = luma_at_offset(caustics, point, np.array([1,0])) lens_normal = np.array([east_luma - west_luma, south_luma - north_luma, 1.0]) lens_normal = lens_normal / np.linalg.norm(lens_normal) refract_vector = refract(np.array([0.0, 0.0, 1.0]), lens_normal, eta) * lens_scale refract_vector = np.round(refract_vector, 3) #print(refract_vector) out_pixel = bilinear_interpolation(image, point+refract_vector[0:2]) out_pixel += (north_luma - south_luma) * lighting_amount out_pixel += (east_luma - west_luma) * lighting_amount return np.clip(out_pixel, 0, 1) noise = np.array([[caustic_noise_kernel(np.array([y,x]), time, size)\ for x in range(self.im_size)] for y in range(self.im_size)]) noise = gaussian_filter(noise, sigma=softening) image = image.astype(np.float32) / 255 out = np.array([[luma_based_refract(np.array([y,x]), image, noise, eta, lens_scale, lighting_amount)\ for x in range(self.im_size)] for y in range(self.im_size)]) return np.clip((out * 255).astype(np.uint8), 0, 255) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - imagenet_c_bar/corrupt.py [732:803]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - softening = 1 return { 'time' : time, 'size' : size, 'eta' : eta, 'lens_scale' : lens_scale, 'lighting_amount': lighting_amount, 'softening' : softening} def transform(self, image, time, size, eta, lens_scale, lighting_amount, softening): def caustic_noise_kernel(point, time, size): point = point / size p = (point % 1) * 6.28318530718 - 250 i = p.copy() c = 1.0 inten = 0.005 for n in range(5): t = time * (1.0 - (3.5 / (n+1))) i = p + np.array([np.cos(t-i[0])+np.sin(t+i[1]),np.sin(t-i[1])+np.cos(t+i[0])]) length = np.sqrt((p[0] / (np.sin(i[0]+t)/inten))**2 + (p[1] / (np.cos(i[1]+t)/inten))**2) c += 1.0/length c /= 5.0 c = 1.17 - c ** 1.4 color = np.clip(np.abs(c) ** 8.0, 0, 1) return np.array([color, color, color]) def refract(incident, normal, eta): if np.abs(np.dot(incident, normal)) >= 1.0 - 1e-3: return incident angle = np.arccos(np.dot(incident, normal)) out_angle = np.arcsin(np.sin(angle) / eta) out_unrotated = np.array([np.cos(out_angle), np.sin(out_angle), 0.0]) spectator_dim = np.cross(incident, normal) spectator_dim /= np.linalg.norm(spectator_dim) orthogonal_dim = np.cross(normal, spectator_dim) rotation_matrix = np.stack((normal, orthogonal_dim, spectator_dim), axis=0) return np.matmul(np.linalg.inv(rotation_matrix), out_unrotated) def luma_at_offset(image, origin, offset): pixel_value = image[origin[0]+offset[0], origin[1]+offset[1], :]\ if origin[0]+offset[0] >= 0 and origin[0]+offset[0] < image.shape[0]\ and origin[1]+offset[1] >= 0 and origin[1]+offset[1] < image.shape[1]\ else np.array([0.0,0.0,0]) return np.dot(pixel_value, np.array([0.2126, 0.7152, 0.0722])) def luma_based_refract(point, image, caustics, eta, lens_scale, lighting_amount): north_luma = luma_at_offset(caustics, point, np.array([0,-1])) south_luma = luma_at_offset(caustics, point, np.array([0, 1])) west_luma = luma_at_offset(caustics, point, np.array([-1, 0])) east_luma = luma_at_offset(caustics, point, np.array([1,0])) lens_normal = np.array([east_luma - west_luma, south_luma - north_luma, 1.0]) lens_normal = lens_normal / np.linalg.norm(lens_normal) refract_vector = refract(np.array([0.0, 0.0, 1.0]), lens_normal, eta) * lens_scale refract_vector = np.round(refract_vector, 3) out_pixel = bilinear_interpolation(image, point+refract_vector[0:2]) out_pixel += (north_luma - south_luma) * lighting_amount out_pixel += (east_luma - west_luma) * lighting_amount return np.clip(out_pixel, 0, 1) noise = np.array([[caustic_noise_kernel(np.array([y,x]), time, size)\ for x in range(self.im_size)] for y in range(self.im_size)]) noise = gaussian_filter(noise, sigma=softening) image = image.astype(np.float32) / 255 out = np.array([[luma_based_refract(np.array([y,x]), image, noise, eta, lens_scale, lighting_amount)\ for x in range(self.im_size)] for y in range(self.im_size)]) return np.clip((out * 255).astype(np.uint8), 0, 255) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -