Spaces:
Running
Running
File size: 8,621 Bytes
c67f469 | 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 | import math
import torch
import torch.nn.functional as F
def _gaussian_blur_heatmaps(heatmaps: torch.Tensor, kernel: int = 11) -> torch.Tensor:
if kernel % 2 == 0:
raise ValueError("kernel must be odd")
sigma = kernel / 6.0
radius = kernel // 2
x = torch.arange(kernel, device=heatmaps.device, dtype=heatmaps.dtype) - radius
g = torch.exp(-(x ** 2) / (2 * sigma * sigma))
g = g / g.sum()
g_x = g.view(1, 1, 1, kernel)
g_y = g.view(1, 1, kernel, 1)
B, N, H, W = heatmaps.shape
# 🔥 FIX HERE
x_in = heatmaps.reshape(B * N, 1, H, W)
x_in = F.pad(x_in, (radius, radius, 0, 0), mode="reflect")
x_in = F.conv2d(x_in, g_x)
x_in = F.pad(x_in, (0, 0, radius, radius), mode="reflect")
x_in = F.conv2d(x_in, g_y)
return x_in.reshape(B, N, H, W)
def heatmaps_to_coords_dark(
heatmaps: torch.Tensor,
blur_kernel: int = 11,
eps: float = 1e-10,
) -> torch.Tensor:
"""
DARK-style decoding with second-order local refinement.
Args:
heatmaps: [B, N, H, W] or [N, H, W]
blur_kernel: Gaussian blur kernel before log
eps: numerical stability for log
Returns:
coords: [B, N, 2] or [N, 2] in heatmap coordinates
"""
squeeze_batch = False
if heatmaps.ndim == 3:
heatmaps = heatmaps.unsqueeze(0)
squeeze_batch = True
if heatmaps.ndim != 4:
raise ValueError(f"Expected [B, N, H, W] or [N, H, W], got {heatmaps.shape}")
B, N, H, W = heatmaps.shape
# Blur then log, as in DARK-style refinement
hm = _gaussian_blur_heatmaps(heatmaps, kernel=blur_kernel)
hm = torch.clamp(hm, min=eps).log()
# Coarse argmax
flat = hm.view(B, N, -1)
idx = flat.argmax(dim=-1)
py = (idx // W).long()
px = (idx % W).long()
coords = torch.stack([px.float(), py.float()], dim=-1)
# Refine using local derivatives of log-heatmap
for b in range(B):
for n in range(N):
x = px[b, n].item()
y = py[b, n].item()
# Need 1-pixel neighborhood for derivatives
if x < 1 or x > W - 2 or y < 1 or y > H - 2:
continue
patch = hm[b, n]
dx = 0.5 * (patch[y, x + 1] - patch[y, x - 1])
dy = 0.5 * (patch[y + 1, x] - patch[y - 1, x])
dxx = patch[y, x + 1] - 2 * patch[y, x] + patch[y, x - 1]
dyy = patch[y + 1, x] - 2 * patch[y, x] + patch[y - 1, x]
dxy = 0.25 * (
patch[y + 1, x + 1]
- patch[y + 1, x - 1]
- patch[y - 1, x + 1]
+ patch[y - 1, x - 1]
)
grad = torch.stack([dx, dy]) # [2]
hessian = torch.stack(
[
torch.stack([dxx, dxy]),
torch.stack([dxy, dyy]),
]
) # [2, 2]
# Solve offset = -H^{-1} g
det = hessian[0, 0] * hessian[1, 1] - hessian[0, 1] * hessian[1, 0]
if torch.abs(det) < 1e-6:
continue
try:
offset = -torch.linalg.solve(hessian, grad)
except RuntimeError:
continue
# Keep refinement bounded; if huge, it's unstable
if torch.all(torch.abs(offset) <= 1.5):
coords[b, n, 0] += offset[0]
coords[b, n, 1] += offset[1]
if squeeze_batch:
coords = coords[0]
return coords
def heatmap_coords_to_image_coords(
coords: torch.Tensor,
image_size: tuple,
heatmap_size: tuple,
) -> torch.Tensor:
"""
Map coordinates from heatmap space back to image space.
Args:
coords: [B, N, 2] or [N, 2]
image_size: (H_img, W_img)
heatmap_size: (H_hm, W_hm)
"""
H_img, W_img = image_size
H_hm, W_hm = heatmap_size
out = coords.clone()
out[..., 0] *= (W_img / W_hm)
out[..., 1] *= (H_img / H_hm)
return out
def gaussian2d(size: int, sigma: float, device=None) -> torch.Tensor:
"""
Create a 2D Gaussian kernel of shape [size, size].
"""
coords = torch.arange(size, device=device, dtype=torch.float32)
center = (size - 1) / 2.0
x = coords - center
y = coords - center
yy, xx = torch.meshgrid(y, x, indexing="ij")
g = torch.exp(-(xx**2 + yy**2) / (2 * sigma * sigma))
return g
def draw_gaussian(
heatmap: torch.Tensor,
center_x: float,
center_y: float,
sigma: float,
) -> torch.Tensor:
"""
Draw a Gaussian on a single heatmap in-place.
Args:
heatmap: [H, W]
center_x, center_y: landmark coordinates in heatmap space
sigma: Gaussian sigma in heatmap pixels
"""
H, W = heatmap.shape
radius = int(3 * sigma)
size = 2 * radius + 1
mu_x = int(round(center_x.item()))
mu_y = int(round(center_y.item()))
left = min(mu_x, radius)
right = min(W - mu_x - 1, radius)
top = min(mu_y, radius)
bottom = min(H - mu_y - 1, radius)
if left < 0 or right < 0 or top < 0 or bottom < 0:
return heatmap
g = gaussian2d(size=size, sigma=sigma, device=heatmap.device)
g_x0 = radius - left
g_x1 = radius + right + 1
g_y0 = radius - top
g_y1 = radius + bottom + 1
h_x0 = mu_x - left
h_x1 = mu_x + right + 1
h_y0 = mu_y - top
h_y1 = mu_y + bottom + 1
heatmap[h_y0:h_y1, h_x0:h_x1] = torch.maximum(
heatmap[h_y0:h_y1, h_x0:h_x1],
g[g_y0:g_y1, g_x0:g_x1],
)
return heatmap
def generate_heatmaps(
landmarks: torch.Tensor,
image_size: tuple,
heatmap_size: tuple,
sigma: float = 2.0,
) -> torch.Tensor:
"""
Generate Gaussian heatmaps for landmark detection.
Args:
landmarks: [N, 2] tensor of (x, y) in original image coordinates
image_size: (H_img, W_img)
heatmap_size: (H_hm, W_hm)
sigma: Gaussian sigma in heatmap pixels
Returns:
heatmaps: [N, H_hm, W_hm]
"""
if landmarks.ndim != 2 or landmarks.shape[1] != 2:
raise ValueError(f"Expected landmarks shape [N, 2], got {landmarks.shape}")
H_img, W_img = image_size
H_hm, W_hm = heatmap_size
scale_x = W_hm / W_img
scale_y = H_hm / H_img
device = landmarks.device
num_landmarks = landmarks.shape[0]
heatmaps = torch.zeros((num_landmarks, H_hm, W_hm), dtype=torch.float32, device=device)
for i in range(num_landmarks):
x, y = landmarks[i]
x_hm = x * scale_x
y_hm = y * scale_y
if 0 <= x_hm < W_hm and 0 <= y_hm < H_hm:
draw_gaussian(heatmaps[i], x_hm, y_hm, sigma=sigma)
return heatmaps
def generate_batch_heatmaps(
landmarks_batch: torch.Tensor,
image_size: tuple,
heatmap_size: tuple,
sigma: float = 2.0,
) -> torch.Tensor:
"""
Batch version.
Args:
landmarks_batch: [B, N, 2]
image_size: (H_img, W_img)
heatmap_size: (H_hm, W_hm)
Returns:
heatmaps: [B, N, H_hm, W_hm]
"""
if landmarks_batch.ndim != 3 or landmarks_batch.shape[-1] != 2:
raise ValueError(f"Expected [B, N, 2], got {landmarks_batch.shape}")
out = []
for b in range(landmarks_batch.shape[0]):
hm = generate_heatmaps(
landmarks=landmarks_batch[b],
image_size=image_size,
heatmap_size=heatmap_size,
sigma=sigma,
)
out.append(hm)
return torch.stack(out, dim=0)
def heatmaps_to_coords_argmax(heatmaps: torch.Tensor) -> torch.Tensor:
"""
Decode coordinates from heatmaps using argmax.
Args:
heatmaps: [B, N, H, W] or [N, H, W]
Returns:
coords: [B, N, 2] or [N, 2] in heatmap coordinates
"""
squeeze_batch = False
if heatmaps.ndim == 3:
heatmaps = heatmaps.unsqueeze(0)
squeeze_batch = True
B, N, H, W = heatmaps.shape
flat = heatmaps.view(B, N, -1)
idx = flat.argmax(dim=-1)
y = idx // W
x = idx % W
coords = torch.stack([x.float(), y.float()], dim=-1)
if squeeze_batch:
coords = coords[0]
return coords
def heatmap_coords_to_image_coords(
coords: torch.Tensor,
image_size: tuple,
heatmap_size: tuple,
) -> torch.Tensor:
"""
Map coordinates from heatmap space back to image space.
"""
H_img, W_img = image_size
H_hm, W_hm = heatmap_size
scale_x = W_img / W_hm
scale_y = H_img / H_hm
out = coords.clone()
out[..., 0] = out[..., 0] * scale_x
out[..., 1] = out[..., 1] * scale_y
return out |