File size: 9,683 Bytes
789eef1 | 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 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import math
import numpy as np
import cv2
def gaussian_blur(heatmaps: np.ndarray, kernel: int = 11) -> np.ndarray:
"""Modulate heatmap distribution with Gaussian.
Note:
- num_keypoints: K
- heatmap height: H
- heatmap width: W
Args:
heatmaps (np.ndarray[K, H, W]): model predicted heatmaps.
kernel (int): Gaussian kernel size (K) for modulation, which should
match the heatmap gaussian sigma when training.
K=17 for sigma=3 and k=11 for sigma=2.
Returns:
np.ndarray ([K, H, W]): Modulated heatmap distribution.
"""
assert kernel % 2 == 1
border = (kernel - 1) // 2
K, H, W = heatmaps.shape
for k in range(K):
origin_max = np.max(heatmaps[k])
dr = np.zeros((H + 2 * border, W + 2 * border), dtype=np.float32)
dr[border:-border, border:-border] = heatmaps[k].copy()
dr = cv2.GaussianBlur(dr, (kernel, kernel), 0)
heatmaps[k] = dr[border:-border, border:-border].copy()
heatmaps[k] *= origin_max / np.max(heatmaps[k])
return heatmaps
def get_heatmap_maximum(heatmaps: np.ndarray):
"""Get maximum response location and value from heatmaps.
Note:
batch_size: B
num_keypoints: K
heatmap height: H
heatmap width: W
Args:
heatmaps (np.ndarray): Heatmaps in shape (K, H, W) or (B, K, H, W)
Returns:
tuple:
- locs (np.ndarray): locations of maximum heatmap responses in shape
(K, 2) or (B, K, 2)
- vals (np.ndarray): values of maximum heatmap responses in shape
(K,) or (B, K)
"""
assert isinstance(heatmaps,
np.ndarray), ('heatmaps should be numpy.ndarray')
assert heatmaps.ndim == 3 or heatmaps.ndim == 4, (
f'Invalid shape {heatmaps.shape}')
if heatmaps.ndim == 3:
K, H, W = heatmaps.shape
B = None
heatmaps_flatten = heatmaps.reshape(K, -1)
else:
B, K, H, W = heatmaps.shape
heatmaps_flatten = heatmaps.reshape(B * K, -1)
y_locs, x_locs = np.unravel_index(
np.argmax(heatmaps_flatten, axis=1), shape=(H, W))
locs = np.stack((x_locs, y_locs), axis=-1).astype(np.float32)
vals = np.amax(heatmaps_flatten, axis=1)
locs[vals <= 0.] = -1
if B:
locs = locs.reshape(B, K, 2)
vals = vals.reshape(B, K)
return locs, vals
def refine_keypoints_dark_udp(keypoints: np.ndarray, heatmaps: np.ndarray,
blur_kernel_size: int) -> np.ndarray:
"""Refine keypoint predictions using distribution aware coordinate decoding
for UDP. See `UDP`_ for details. The operation is in-place.
Note:
- instance number: N
- keypoint number: K
- keypoint dimension: D
- heatmap size: [W, H]
Args:
keypoints (np.ndarray): The keypoint coordinates in shape (N, K, D)
heatmaps (np.ndarray): The heatmaps in shape (K, H, W)
blur_kernel_size (int): The Gaussian blur kernel size of the heatmap
modulation
Returns:
np.ndarray: Refine keypoint coordinates in shape (N, K, D)
.. _`UDP`: https://arxiv.org/abs/1911.07524
"""
N, K = keypoints.shape[:2]
H, W = heatmaps.shape[1:]
# modulate heatmaps
heatmaps = gaussian_blur(heatmaps, blur_kernel_size)
np.clip(heatmaps, 1e-3, 50., heatmaps)
np.log(heatmaps, heatmaps)
heatmaps_pad = np.pad(
heatmaps, ((0, 0), (1, 1), (1, 1)), mode='edge').flatten()
for n in range(N):
index = keypoints[n, :, 0] + 1 + (keypoints[n, :, 1] + 1) * (W + 2)
index += (W + 2) * (H + 2) * np.arange(0, K)
index = index.astype(int).reshape(-1, 1)
i_ = heatmaps_pad[index]
ix1 = heatmaps_pad[index + 1]
iy1 = heatmaps_pad[index + W + 2]
ix1y1 = heatmaps_pad[index + W + 3]
ix1_y1_ = heatmaps_pad[index - W - 3]
ix1_ = heatmaps_pad[index - 1]
iy1_ = heatmaps_pad[index - 2 - W]
dx = 0.5 * (ix1 - ix1_)
dy = 0.5 * (iy1 - iy1_)
derivative = np.concatenate([dx, dy], axis=1)
derivative = derivative.reshape(K, 2, 1)
dxx = ix1 - 2 * i_ + ix1_
dyy = iy1 - 2 * i_ + iy1_
dxy = 0.5 * (ix1y1 - ix1 - iy1 + i_ + i_ - ix1_ - iy1_ + ix1_y1_)
hessian = np.concatenate([dxx, dxy, dxy, dyy], axis=1)
hessian = hessian.reshape(K, 2, 2)
hessian = np.linalg.inv(hessian + np.finfo(np.float32).eps * np.eye(2))
keypoints[n] -= np.einsum('imn,ink->imk', hessian,
derivative).squeeze()
return keypoints
def udp_decode(heatmaps, input_size, heatmap_size, blur_kernel_size=11) -> np.ndarray:
"""UDP decoding for keypoint location refinement.
Note:
- num_keypoints: K
- heatmap height: H
- heatmap width: W
Args:
heatmaps (np.ndarray[K, H, W]): model predicted heatmaps.
blur_kernel_size (int): Gaussian kernel size (K) for modulation, which
should match the heatmap gaussian sigma when training.
K=17 for sigma=3 and k=11 for sigma=2.
Returns:
np.ndarray ([K, H, W]): Refined keypoint locations.
"""
keypoints, scores = get_heatmap_maximum(heatmaps)
# unsqueeze the instance dimension for single-instance results
keypoints = keypoints[None]
scores = scores[None]
keypoints = refine_keypoints_dark_udp(
keypoints, heatmaps, blur_kernel_size=blur_kernel_size)
W, H = heatmap_size
keypoints = (keypoints / [W - 1, H - 1]) * input_size
return keypoints, scores
def get_udp_warp_matrix(
center: np.ndarray,
scale: np.ndarray,
rot: float,
output_size,
) -> np.ndarray:
"""Calculate the affine transformation matrix under the unbiased
constraint. See `UDP (CVPR 2020)`_ for details.
Note:
- The bbox number: N
Args:
center (np.ndarray[2, ]): Center of the bounding box (x, y).
scale (np.ndarray[2, ]): Scale of the bounding box
wrt [width, height].
rot (float): Rotation angle (degree).
output_size (tuple): Size ([w, h]) of the output image
Returns:
np.ndarray: A 2x3 transformation matrix
.. _`UDP (CVPR 2020)`: https://arxiv.org/abs/1911.07524
"""
assert len(center) == 2
assert len(scale) == 2
assert len(output_size) == 2
input_size = center * 2
rot_rad = np.deg2rad(rot)
warp_mat = np.zeros((2, 3), dtype=np.float32)
scale_x = (output_size[0] - 1) / scale[0]
scale_y = (output_size[1] - 1) / scale[1]
warp_mat[0, 0] = math.cos(rot_rad) * scale_x
warp_mat[0, 1] = -math.sin(rot_rad) * scale_x
warp_mat[0, 2] = scale_x * (-0.5 * input_size[0] * math.cos(rot_rad) +
0.5 * input_size[1] * math.sin(rot_rad) +
0.5 * scale[0])
warp_mat[1, 0] = math.sin(rot_rad) * scale_y
warp_mat[1, 1] = math.cos(rot_rad) * scale_y
warp_mat[1, 2] = scale_y * (-0.5 * input_size[0] * math.sin(rot_rad) -
0.5 * input_size[1] * math.cos(rot_rad) +
0.5 * scale[1])
return warp_mat
def top_down_affine_transform(img, bbox, padding=1.25):
"""
Args:
img (np.ndarray): Image to be transformed.
bbox (np.ndarray): Bounding box to be transformed.
padding (int): Padding size.
Returns:
np.ndarray: Transformed image.
np.ndarray: Transformed bounding box.
"""
dim = bbox.ndim
if dim == 1:
bbox = bbox[None, :]
x1, y1, x2, y2 = np.hsplit(bbox, [1, 2, 3])
center = np.hstack([x1 + x2, y1 + y2]) * 0.5
scale = np.hstack([x2 - x1, y2 - y1]) * padding
if dim == 1:
center = center[0]
scale = scale[0]
h, w = img.shape[:2]
warp_size = (int(w), int(h))
aspect_ratio = w / h
# reshape bbox to fixed aspect ratio
box_w, box_h = np.hsplit(scale, [1])
scale = np.where(box_w > box_h * aspect_ratio,
np.hstack([box_w, box_w / aspect_ratio]),
np.hstack([box_h * aspect_ratio, box_h]))
rot = 0.
warp_mat = get_udp_warp_matrix(
center, scale, rot, output_size=(w, h))
img = cv2.warpAffine(
img, warp_mat, warp_size, flags=cv2.INTER_LINEAR)
return img, [center], [scale]
def nms(dets: np.ndarray, thr: float):
"""Greedily select boxes with high confidence and overlap <= thr.
Args:
dets (np.ndarray): [[x1, y1, x2, y2, score]].
thr (float): Retain overlap < thr.
Returns:
list: Indexes to keep.
"""
if len(dets) == 0:
return []
x1 = dets[:, 0]
y1 = dets[:, 1]
x2 = dets[:, 2]
y2 = dets[:, 3]
scores = dets[:, 4]
areas = (x2 - x1 + 1) * (y2 - y1 + 1)
order = scores.argsort()[::-1]
keep = []
while len(order) > 0:
i = order[0]
keep.append(i)
xx1 = np.maximum(x1[i], x1[order[1:]])
yy1 = np.maximum(y1[i], y1[order[1:]])
xx2 = np.minimum(x2[i], x2[order[1:]])
yy2 = np.minimum(y2[i], y2[order[1:]])
w = np.maximum(0.0, xx2 - xx1 + 1)
h = np.maximum(0.0, yy2 - yy1 + 1)
inter = w * h
ovr = inter / (areas[i] + areas[order[1:]] - inter)
inds = np.where(ovr <= thr)[0]
order = order[inds + 1]
return keep
|