File size: 12,668 Bytes
5732928 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 |
import cv2
import torch
import random
import numpy as np
import torch.nn.functional as F
from lib.core.config import cfg
from lib.utils.human_models import mano
def get_aug_config_contact():
# Augmentation intensity factors
scale_factor = 0.25
rot_factor = 30
color_factor = 0.2
trans_factor = 0.1 # Translation range (recommended 0.1 to 0.2)
noise_std = 0.02 # Gaussian noise strength
motion_blur_prob = 0.15 # Probability of applying motion blur
extreme_crop_prob = 0.1 # Probability for extreme cropping
extreme_crop_lvl = 0.3 # Crop intensity (recommended 0.2 to 0.4)
low_res_prob = 0.05 # Probability for applying low resolution
low_res_scale_range = (0.15, 0.5) # Range for low-res scaling
# Scaling augmentation
scale = np.clip(np.random.randn(), -1.0, 1.0) * scale_factor + 1.0
# Rotation augmentation
rot = np.clip(np.random.randn(), -2.0, 2.0) * rot_factor if random.random() <= 0.6 else 0
# Color augmentation
c_up = 1.0 + color_factor
c_low = 1.0 - color_factor
color_scale = np.array([
random.uniform(c_low, c_up),
random.uniform(c_low, c_up),
random.uniform(c_low, c_up)
])
# Flipping augmentation
do_flip = random.random() <= 0.5
# Translation augmentation
tx = np.clip(np.random.randn(), -1.0, 1.0) * trans_factor
ty = np.clip(np.random.randn(), -1.0, 1.0) * trans_factor
# Extreme cropping augmentation
do_extreme_crop = random.random() <= extreme_crop_prob
# Noise augmentation (returns standard deviation for Gaussian noise injection)
add_noise = random.random() <= 0.3 # 30% chance of adding noise
noise_std = noise_std if add_noise else 0.0
# Motion blur augmentation
apply_motion_blur = random.random() <= motion_blur_prob
motion_blur_kernel_size = random.choice([3, 5, 7]) if apply_motion_blur else 0
# Low-resolution augmentation
apply_low_res = random.random() <= low_res_prob
low_res_scale = random.uniform(*low_res_scale_range) if apply_low_res else 1.0
return {
'scale': scale,
'rot': rot,
'color_scale': color_scale,
'do_flip': do_flip,
'tx': tx,
'ty': ty,
'do_extreme_crop': do_extreme_crop,
'extreme_crop_lvl': extreme_crop_lvl if do_extreme_crop else 0,
'noise_std': noise_std,
'motion_blur_kernel_size': motion_blur_kernel_size,
'low_res_scale': low_res_scale # Added low-res scale parameter
}
def rotate_2d(pt_2d, rot_rad):
x = pt_2d[0]
y = pt_2d[1]
sn, cs = np.sin(rot_rad), np.cos(rot_rad)
xx = x * cs - y * sn
yy = x * sn + y * cs
return np.array([xx, yy], dtype=np.float32)
def gen_trans_from_patch_cv(c_x, c_y, src_width, src_height, dst_width, dst_height, scale, rot, inv=False):
# augment size with scale
src_w = src_width * scale
src_h = src_height * scale
src_center = np.array([c_x, c_y], dtype=np.float32)
# augment rotation
rot_rad = np.pi * rot / 180
src_downdir = rotate_2d(np.array([0, src_h * 0.5], dtype=np.float32), rot_rad)
src_rightdir = rotate_2d(np.array([src_w * 0.5, 0], dtype=np.float32), rot_rad)
dst_w = dst_width
dst_h = dst_height
dst_center = np.array([dst_w * 0.5, dst_h * 0.5], dtype=np.float32)
dst_downdir = np.array([0, dst_h * 0.5], dtype=np.float32)
dst_rightdir = np.array([dst_w * 0.5, 0], dtype=np.float32)
src = np.zeros((3, 2), dtype=np.float32)
src[0, :] = src_center
src[1, :] = src_center + src_downdir
src[2, :] = src_center + src_rightdir
dst = np.zeros((3, 2), dtype=np.float32)
dst[0, :] = dst_center
dst[1, :] = dst_center + dst_downdir
dst[2, :] = dst_center + dst_rightdir
if inv:
trans = cv2.getAffineTransform(np.float32(dst), np.float32(src))
else:
trans = cv2.getAffineTransform(np.float32(src), np.float32(dst))
trans = trans.astype(np.float32)
return trans
def generate_patch_image_contact(cvimg, bbox, scale, rot, do_flip, out_shape, tx=0.0, ty=0.0, bkg_color='black'):
img = cvimg.copy()
img_height, img_width, img_channels = img.shape
bb_c_x = float(bbox[0] + 0.5 * bbox[2])
bb_c_y = float(bbox[1] + 0.5 * bbox[3])
bb_width = float(bbox[2])
bb_height = float(bbox[3])
if bkg_color == 'white':
borderMode=cv2.BORDER_CONSTANT
borderValue=(255, 255, 255)
else:
borderMode=cv2.BORDER_CONSTANT
borderValue=(0, 0, 0)
if do_flip:
img = img[:, ::-1, :]
bb_c_x = img_width - bb_c_x - 1
# Add translation offset
bb_c_x += tx * img_width
bb_c_y += ty * img_height
trans = gen_trans_from_patch_cv(bb_c_x, bb_c_y, bb_width, bb_height,
out_shape[1], out_shape[0], scale, rot)
img_patch = cv2.warpAffine(img, trans, (int(out_shape[1]), int(out_shape[0])), flags=cv2.INTER_LINEAR, borderMode=borderMode, borderValue=borderValue)
img_patch = img_patch.astype(np.float32)
inv_trans = gen_trans_from_patch_cv(bb_c_x, bb_c_y, bb_width, bb_height,
out_shape[1], out_shape[0], scale, rot, inv=True)
return img_patch, trans, inv_trans
def augmentation_contact(img, bbox, data_split, enforce_flip=None, bkg_color='black'):
if data_split == 'train':
aug_params = get_aug_config_contact()
else:
aug_params = {
'scale': 1.0,
'rot': 0.0,
'color_scale': np.array([1, 1, 1]),
'do_flip': False,
'tx': 0.0,
'ty': 0.0,
'do_extreme_crop': False,
'extreme_crop_lvl': 0.0,
'noise_std': 0.0,
'motion_blur_kernel_size': 0,
'low_res_scale': 1.0 # No low-res in non-training mode
}
# Enforce flip if specified
if enforce_flip is not None:
aug_params['do_flip'] = enforce_flip
# Apply geometric augmentations (scaling, rotation, flipping)
img, trans, inv_trans = generate_patch_image_contact(
img, bbox, aug_params['scale'], aug_params['rot'],
aug_params['do_flip'], cfg.MODEL.input_img_shape,
aug_params['tx'], aug_params['ty'], bkg_color
)
# Apply low-resolution augmentation
if aug_params['low_res_scale'] < 1.0: # Only apply if scaling down
img = apply_low_res(img, aug_params['low_res_scale'])
# Apply color augmentation
img = np.clip(img * aug_params['color_scale'][None, None, :], 0, 255)
# Apply extreme cropping
if aug_params['do_extreme_crop']:
img = apply_extreme_crop(img, aug_params['extreme_crop_lvl'])
# Apply noise augmentation
if aug_params['noise_std'] > 0:
img = add_gaussian_noise(img, aug_params['noise_std'])
# Apply motion blur augmentation
if aug_params['motion_blur_kernel_size'] > 0:
img = apply_motion_blur(img, aug_params['motion_blur_kernel_size'])
return img, trans, inv_trans, aug_params['rot'], aug_params['do_flip'], aug_params['color_scale']
def apply_extreme_crop(img, crop_lvl):
"""Extreme cropping: Aggressively crop the image."""
h, w = img.shape[:2]
crop_size = max(1, int(min(h, w) * (1 - crop_lvl))) # Prevent zero-size crops
start_x = random.randint(0, max(0, w - crop_size))
start_y = random.randint(0, max(0, h - crop_size))
cropped_img = img[start_y:start_y + crop_size, start_x:start_x + crop_size]
# Preserve aspect ratio during resizing
return cv2.resize(cropped_img, (w, h), interpolation=cv2.INTER_LINEAR)
def add_gaussian_noise(img, noise_std):
"""Add Gaussian noise to the image with proper scaling for data type."""
noise = np.random.normal(0, noise_std, img.shape).astype(np.float32)
if img.dtype == np.uint8:
noisy_img = np.clip(img + noise * 255, 0, 255).astype(np.uint8)
elif img.dtype == np.float32:
noisy_img = np.clip(img + noise, 0.0, 1.0).astype(np.float32)
elif img.dtype == np.float64:
noisy_img = np.clip(img + noise, 0.0, 1.0).astype(np.float64)
else:
raise TypeError("Unsupported image dtype. Expected uint8 or float32.")
return noisy_img
def apply_motion_blur(img, kernel_size):
"""Apply motion blur to the image with a random direction."""
kernel = np.zeros((kernel_size, kernel_size))
direction = random.choice(['horizontal', 'vertical', 'diagonal'])
if direction == 'horizontal':
kernel[(kernel_size - 1) // 2, :] = np.ones(kernel_size)
elif direction == 'vertical':
kernel[:, (kernel_size - 1) // 2] = np.ones(kernel_size)
elif direction == 'diagonal':
np.fill_diagonal(kernel, 1)
kernel /= kernel_size # Normalize the kernel
return cv2.filter2D(img, -1, kernel, borderType=cv2.BORDER_REFLECT)
def apply_low_res(img, scale_factor=0.25):
"""Simulate low-resolution effect by downsampling and upsampling."""
if not (0 < scale_factor < 1):
raise ValueError("scale_factor should be between 0 and 1.")
h, w = img.shape[:2]
# Calculate target dimensions for downsampling
downsampled_size = (max(1, int(w * scale_factor)), max(1, int(h * scale_factor)))
# Downsample using INTER_AREA for better quality in aggressive downsampling
low_res_img = cv2.resize(img, downsampled_size, interpolation=cv2.INTER_AREA)
# Upsample using INTER_NEAREST for strong pixelation effect
return cv2.resize(low_res_img, (w, h), interpolation=cv2.INTER_NEAREST).astype(img.dtype)
def process_human_model_output_orig(human_model_param, cam_param):
pose, shape, trans = human_model_param['pose'], human_model_param['shape'], human_model_param['trans']
hand_type = human_model_param['hand_type']
trans = human_model_param['trans']
pose = torch.FloatTensor(pose).view(-1,3); shape = torch.FloatTensor(shape).view(1,-1); # mano parameters (pose: 48 dimension, shape: 10 dimension)
trans = torch.FloatTensor(trans).view(1,-1) # translation vector
# apply camera extrinsic (rotation)
# merge root pose and camera rotation
if 'R' in cam_param:
R = np.array(cam_param['R'], dtype=np.float32).reshape(3,3)
root_pose = pose[mano.orig_root_joint_idx,:].numpy()
root_pose, _ = cv2.Rodrigues(root_pose)
root_pose, _ = cv2.Rodrigues(np.dot(R,root_pose))
pose[mano.orig_root_joint_idx] = torch.from_numpy(root_pose).view(3)
# get root joint coordinate
root_pose = pose[mano.orig_root_joint_idx].view(1,3)
hand_pose = torch.cat((pose[:mano.orig_root_joint_idx,:], pose[mano.orig_root_joint_idx+1:,:])).view(1,-1)
with torch.no_grad():
output = mano.layer[hand_type](betas=shape, hand_pose=hand_pose, global_orient=root_pose, transl=trans)
mesh_coord = output.vertices[0].numpy()
joint_coord = np.dot(mano.joint_regressor, mesh_coord)
# apply camera exrinsic (translation)
# compenstate rotation (translation from origin to root joint was not cancled)
if 'R' in cam_param and 't' in cam_param:
R, t = np.array(cam_param['R'], dtype=np.float32).reshape(3,3), np.array(cam_param['t'], dtype=np.float32).reshape(1,3)
root_coord = joint_coord[mano.root_joint_idx,None,:]
joint_coord = joint_coord - root_coord + np.dot(R, root_coord.transpose(1,0)).transpose(1,0) + t
mesh_coord = mesh_coord - root_coord + np.dot(R, root_coord.transpose(1,0)).transpose(1,0) + t
joint_cam_orig = joint_coord.copy()
mesh_cam_orig = mesh_coord.copy()
pose_orig, shape_orig, trans_orig = torch.cat((root_pose, hand_pose), dim=-1)[0].detach().cpu().numpy(), shape[0].detach().cpu().numpy(), trans[0].detach().cpu().numpy()
return mesh_cam_orig, joint_cam_orig, pose_orig, shape_orig, trans_orig
def mask2bbox(mask, expansion_factor=1.0):
# Find non-zero elements (object pixels)
coords = np.argwhere(mask)
# Extract bounding box coordinates
y_min, x_min = coords.min(axis=0)
y_max, x_max = coords.max(axis=0)
# Compute width and height
width = x_max - x_min + 1
height = y_max - y_min + 1
# Expand bounding box
if expansion_factor > 0:
x_min = max(0, int(x_min - width * expansion_factor / 2))
y_min = max(0, int(y_min - height * expansion_factor / 2))
x_max = min(mask.shape[1] - 1, int(x_max + width * expansion_factor / 2))
y_max = min(mask.shape[0] - 1, int(y_max + height * expansion_factor / 2))
# Recalculate width and height after expansion
width = x_max - x_min + 1
height = y_max - y_min + 1
return (x_min, y_min, width, height) |