Spaces:
Running
Running
File size: 10,931 Bytes
433e26f 6071b18 433e26f 6071b18 433e26f 6071b18 433e26f 6071b18 433e26f 6071b18 433e26f 6071b18 433e26f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 | """Training data augmentation pipeline for LandmarkDiff.
Provides domain-specific augmentations that maintain landmark consistency:
- Geometric: flip, rotation, affine (landmarks co-transformed)
- Photometric: color jitter, brightness, contrast (applied to images only)
- Skin-tone augmentation: ITA-space perturbation for Fitzpatrick balance
- Conditioning augmentation: noise injection, dropout for robustness
All augmentations preserve the correspondence between:
input_image ↔ conditioning_image ↔ target_image ↔ mask
"""
from __future__ import annotations
from dataclasses import dataclass
import cv2
import numpy as np
@dataclass
class AugmentationConfig:
"""Augmentation parameters."""
# Geometric
random_flip: bool = True
random_rotation_deg: float = 5.0
random_scale: tuple[float, float] = (0.95, 1.05)
random_translate: float = 0.02 # fraction of image size
# Photometric (images only, not conditioning)
brightness_range: tuple[float, float] = (0.9, 1.1)
contrast_range: tuple[float, float] = (0.9, 1.1)
saturation_range: tuple[float, float] = (0.9, 1.1)
hue_shift_range: float = 5.0 # degrees
# Conditioning augmentation
conditioning_dropout_prob: float = 0.1
conditioning_noise_std: float = 0.02
# Skin-tone augmentation
ita_perturbation_std: float = 3.0 # ITA angle noise
seed: int | None = None
def augment_training_sample(
input_image: np.ndarray,
target_image: np.ndarray,
conditioning: np.ndarray,
mask: np.ndarray,
landmarks_src: np.ndarray | None = None,
landmarks_dst: np.ndarray | None = None,
config: AugmentationConfig | None = None,
rng: np.random.Generator | None = None,
) -> dict[str, np.ndarray]:
"""Apply consistent augmentations to a training sample.
All spatial transforms are applied to images AND landmarks together
so correspondence is preserved.
Args:
input_image: (H, W, 3) original face image (uint8 BGR).
target_image: (H, W, 3) target face image (uint8 BGR).
conditioning: (H, W, 3) conditioning image (uint8).
mask: (H, W) or (H, W, 1) float32 mask.
landmarks_src: (N, 2) normalized [0,1] source landmark coords.
landmarks_dst: (N, 2) normalized [0,1] target landmark coords.
config: Augmentation parameters.
rng: Random generator for reproducibility.
Returns:
Dict with augmented versions of all inputs.
"""
if config is None:
config = AugmentationConfig()
if rng is None:
rng = np.random.default_rng(config.seed)
h, w = input_image.shape[:2]
out_input = input_image.copy()
out_target = target_image.copy()
out_cond = conditioning.copy()
out_mask = mask.copy()
out_lm_src = landmarks_src.copy() if landmarks_src is not None else None
out_lm_dst = landmarks_dst.copy() if landmarks_dst is not None else None
# --- Geometric augmentations (applied to all) ---
# Random horizontal flip
if config.random_flip and rng.random() < 0.5:
out_input = np.ascontiguousarray(out_input[:, ::-1])
out_target = np.ascontiguousarray(out_target[:, ::-1])
out_cond = np.ascontiguousarray(out_cond[:, ::-1])
out_mask = np.ascontiguousarray(out_mask[:, ::-1] if out_mask.ndim == 2
else out_mask[:, ::-1, :])
if out_lm_src is not None:
out_lm_src[:, 0] = 1.0 - out_lm_src[:, 0]
if out_lm_dst is not None:
out_lm_dst[:, 0] = 1.0 - out_lm_dst[:, 0]
# Random rotation + scale + translate
if config.random_rotation_deg > 0 or config.random_scale != (1.0, 1.0):
angle = rng.uniform(-config.random_rotation_deg, config.random_rotation_deg)
scale = rng.uniform(config.random_scale[0], config.random_scale[1])
tx = rng.uniform(-config.random_translate, config.random_translate) * w
ty = rng.uniform(-config.random_translate, config.random_translate) * h
center = (w / 2, h / 2)
M = cv2.getRotationMatrix2D(center, angle, scale)
M[0, 2] += tx
M[1, 2] += ty
out_input = cv2.warpAffine(out_input, M, (w, h),
borderMode=cv2.BORDER_REFLECT_101)
out_target = cv2.warpAffine(out_target, M, (w, h),
borderMode=cv2.BORDER_REFLECT_101)
out_cond = cv2.warpAffine(out_cond, M, (w, h),
borderMode=cv2.BORDER_CONSTANT, borderValue=0)
mask_2d = out_mask if out_mask.ndim == 2 else out_mask[:, :, 0]
mask_2d = cv2.warpAffine(mask_2d, M, (w, h),
borderMode=cv2.BORDER_CONSTANT, borderValue=0)
out_mask = mask_2d if out_mask.ndim == 2 else mask_2d[:, :, np.newaxis]
# Transform landmarks
if out_lm_src is not None:
out_lm_src = _transform_landmarks(out_lm_src, M, w, h)
if out_lm_dst is not None:
out_lm_dst = _transform_landmarks(out_lm_dst, M, w, h)
# --- Photometric augmentations (images only, not conditioning/mask) ---
# Brightness
b_factor = rng.uniform(config.brightness_range[0], config.brightness_range[1])
out_input = np.clip(out_input.astype(np.float32) * b_factor, 0, 255).astype(np.uint8)
out_target = np.clip(out_target.astype(np.float32) * b_factor, 0, 255).astype(np.uint8)
# Contrast
c_factor = rng.uniform(config.contrast_range[0], config.contrast_range[1])
mean_in = out_input.mean()
mean_tgt = out_target.mean()
out_input = np.clip(
(out_input.astype(np.float32) - mean_in) * c_factor + mean_in, 0, 255
).astype(np.uint8)
out_target = np.clip(
(out_target.astype(np.float32) - mean_tgt) * c_factor + mean_tgt, 0, 255
).astype(np.uint8)
# Saturation (in HSV space)
s_factor = rng.uniform(config.saturation_range[0], config.saturation_range[1])
if abs(s_factor - 1.0) > 1e-4:
out_input = _adjust_saturation(out_input, s_factor)
out_target = _adjust_saturation(out_target, s_factor)
# Hue shift
if config.hue_shift_range > 0:
hue_delta = rng.uniform(-config.hue_shift_range, config.hue_shift_range)
if abs(hue_delta) > 0.1:
out_input = _shift_hue(out_input, hue_delta)
out_target = _shift_hue(out_target, hue_delta)
# --- Conditioning augmentation ---
# Conditioning dropout (replace with zeros to learn unconditional)
if config.conditioning_dropout_prob > 0 and rng.random() < config.conditioning_dropout_prob:
out_cond = np.zeros_like(out_cond)
# Conditioning noise
if config.conditioning_noise_std > 0:
noise = rng.normal(0, config.conditioning_noise_std * 255, out_cond.shape)
out_cond = np.clip(out_cond.astype(np.float32) + noise, 0, 255).astype(np.uint8)
result = {
"input_image": out_input,
"target_image": out_target,
"conditioning": out_cond,
"mask": out_mask,
}
if out_lm_src is not None:
result["landmarks_src"] = out_lm_src
if out_lm_dst is not None:
result["landmarks_dst"] = out_lm_dst
return result
def _transform_landmarks(
landmarks: np.ndarray, M: np.ndarray, w: int, h: int
) -> np.ndarray:
"""Transform normalized landmarks with an affine matrix."""
# Convert to pixel coords
px = landmarks.copy()
px[:, 0] *= w
px[:, 1] *= h
# Apply affine transform
ones = np.ones((px.shape[0], 1))
px_h = np.hstack([px, ones]) # (N, 3)
transformed = (M @ px_h.T).T # (N, 2)
# Back to normalized
transformed[:, 0] /= w
transformed[:, 1] /= h
return np.clip(transformed, 0.0, 1.0)
def _adjust_saturation(img: np.ndarray, factor: float) -> np.ndarray:
"""Adjust saturation of a BGR image."""
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype(np.float32)
hsv[:, :, 1] = np.clip(hsv[:, :, 1] * factor, 0, 255)
return cv2.cvtColor(hsv.astype(np.uint8), cv2.COLOR_HSV2BGR)
def _shift_hue(img: np.ndarray, delta_deg: float) -> np.ndarray:
"""Shift hue of a BGR image by delta degrees."""
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype(np.float32)
# OpenCV hue range is [0, 180]
hsv[:, :, 0] = (hsv[:, :, 0] + delta_deg / 2) % 180
return cv2.cvtColor(hsv.astype(np.uint8), cv2.COLOR_HSV2BGR)
def augment_skin_tone(
image: np.ndarray,
ita_delta: float = 0.0,
) -> np.ndarray:
"""Augment skin tone by shifting in L*a*b* space.
This helps balance Fitzpatrick representation in training by
simulating different skin tones from existing samples.
Args:
image: (H, W, 3) BGR uint8 image.
ita_delta: ITA angle shift (positive = lighter, negative = darker).
Returns:
Augmented image with shifted skin tone.
"""
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB).astype(np.float32)
# Shift L channel (lightness) based on ITA delta
# ITA = arctan((L-50)/b), so shifting ITA shifts L
l_shift = ita_delta * 0.5 # approximate mapping
lab[:, :, 0] = np.clip(lab[:, :, 0] + l_shift, 0, 255)
# Slightly shift b channel too for more natural tone changes
b_shift = -ita_delta * 0.15
lab[:, :, 2] = np.clip(lab[:, :, 2] + b_shift, 0, 255)
return cv2.cvtColor(lab.astype(np.uint8), cv2.COLOR_LAB2BGR)
class FitzpatrickBalancer:
"""Oversample underrepresented Fitzpatrick types during training.
Maintains per-type counts and generates sampling weights to ensure
equitable training across all skin types.
"""
def __init__(self, target_distribution: dict[str, float] | None = None):
"""Initialize balancer.
Args:
target_distribution: Target fraction per type. Defaults to uniform.
"""
self.target = target_distribution or {
"I": 1/6, "II": 1/6, "III": 1/6,
"IV": 1/6, "V": 1/6, "VI": 1/6,
}
self._counts: dict[str, int] = {}
def register_sample(self, fitz_type: str) -> None:
"""Register a sample's Fitzpatrick type."""
self._counts[fitz_type] = self._counts.get(fitz_type, 0) + 1
def get_sampling_weights(self, fitz_types: list[str]) -> np.ndarray:
"""Compute sampling weights for a list of samples.
Returns weights inversely proportional to type frequency,
so underrepresented types get upsampled.
"""
total = sum(self._counts.values()) or 1
weights = []
for ft in fitz_types:
count = self._counts.get(ft, 1)
freq = count / total
target_freq = self.target.get(ft, 1/6)
# Weight = target / actual (capped for stability)
w = min(target_freq / max(freq, 1e-6), 5.0)
weights.append(w)
w = np.array(weights, dtype=np.float64)
return w / w.sum() # normalize to probability distribution
|