File size: 7,117 Bytes
e1b0735 | 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 | """Visualization and evaluation helpers used by the inference entrypoints."""
import matplotlib
matplotlib.use("Agg") # headless backend, no display needed
import cv2
import matplotlib.pyplot as plt
import numpy as np
from sklearn.metrics import confusion_matrix
def get_cmap(name, n):
"""Discrete colormap, compatible with matplotlib >= 3.9.
``plt.cm.get_cmap`` was removed in 3.9; ``plt.get_cmap(...).resampled(n)``
is the supported spelling, with a fallback for older releases.
"""
cmap = plt.get_cmap(name)
return cmap.resampled(n) if hasattr(cmap, "resampled") else plt.cm.get_cmap(name, n)
def quick_resize(img, max_dim=1024, is_mask=False):
"""Shrink so the longest edge is <= max_dim, preserving aspect ratio.
Masks use nearest-neighbour so class indices survive; images/probability
maps use area interpolation.
"""
h, w = img.shape[:2]
scale = max_dim / float(max(h, w))
if scale < 1.0: # only shrink, never upscale
new_w, new_h = int(w * scale), int(h * scale)
interp = cv2.INTER_NEAREST if is_mask else cv2.INTER_AREA
img = cv2.resize(img, (new_w, new_h), interpolation=interp)
return img
def vis_img(image, pred_mask, foreground_probs_all, template_all, save_path, max_dim=1024):
"""One row per class (image | probability map) plus a final merged-mask row."""
img_small = quick_resize(image, max_dim=max_dim)
pred_mask_small = quick_resize(pred_mask, max_dim=max_dim, is_mask=True)
probs_small = [quick_resize(p, max_dim=max_dim) for p in foreground_probs_all]
num_rows = len(probs_small) + 1
fig, axs = plt.subplots(num_rows, 2, figsize=(12, 4 * num_rows), squeeze=False)
for i in range(num_rows - 1):
axs[i, 0].imshow(img_small)
axs[i, 0].set_title(template_all[i])
axs[i, 0].axis("off")
axs[i, 1].imshow(probs_small[i], cmap="gray", vmin=0, vmax=1)
axs[i, 1].set_title("Predicted Probability")
axs[i, 1].axis("off")
n_classes = len(template_all) + 1
cmap = get_cmap("tab20", n_classes)
axs[num_rows - 1, 0].imshow(img_small)
axs[num_rows - 1, 0].set_title("Raw Image")
axs[num_rows - 1, 0].axis("off")
im1 = axs[num_rows - 1, 1].imshow(pred_mask_small, cmap=cmap, vmin=0,
vmax=n_classes - 1, interpolation="nearest")
axs[num_rows - 1, 1].set_title("Predicted Mask")
axs[num_rows - 1, 1].axis("off")
cbar = fig.colorbar(im1, ax=axs[num_rows - 1, 1], orientation="vertical",
fraction=0.02, pad=0.04)
cbar.set_ticks(range(n_classes))
cbar.set_ticklabels(["background"] + list(template_all))
fig.savefig(save_path, bbox_inches="tight", dpi=200)
plt.close(fig)
def vis_img_bbx(image, pred_mask, true_mask, foreground_probs_all, gt_masks_all,
template_all, save_path, max_dim=1024):
"""One row per class (image | prediction | ground truth) plus a merged row."""
img_small = quick_resize(image, max_dim=max_dim)
pred_mask_small = quick_resize(pred_mask, max_dim=max_dim, is_mask=True)
true_mask_small = quick_resize(true_mask, max_dim=max_dim, is_mask=True)
probs_small = [quick_resize(p, max_dim=max_dim) for p in foreground_probs_all]
gts_small = [quick_resize(g, max_dim=max_dim, is_mask=True) for g in gt_masks_all]
num_rows = len(probs_small) + 1
fig, axs = plt.subplots(num_rows, 3, figsize=(12, 4 * num_rows), squeeze=False)
for i in range(num_rows - 1):
axs[i, 0].imshow(img_small)
axs[i, 0].set_title(template_all[i])
axs[i, 0].axis("off")
axs[i, 1].imshow(probs_small[i], cmap="gray", vmin=0, vmax=1)
axs[i, 1].set_title("Predicted Probability")
axs[i, 1].axis("off")
axs[i, 2].imshow(gts_small[i], cmap="gray", vmin=0, vmax=1)
axs[i, 2].set_title("Ground Truth Mask")
axs[i, 2].axis("off")
n_classes = len(template_all) + 1
cmap = get_cmap("tab20", n_classes)
axs[num_rows - 1, 0].imshow(img_small)
axs[num_rows - 1, 0].set_title("Raw Image")
axs[num_rows - 1, 0].axis("off")
axs[num_rows - 1, 1].imshow(pred_mask_small, cmap=cmap, vmin=0,
vmax=n_classes - 1, interpolation="nearest")
axs[num_rows - 1, 1].set_title("Predicted Mask")
axs[num_rows - 1, 1].axis("off")
im1 = axs[num_rows - 1, 2].imshow(true_mask_small, cmap=cmap, vmin=0,
vmax=n_classes - 1, interpolation="nearest")
axs[num_rows - 1, 2].set_title("Ground Truth Mask")
axs[num_rows - 1, 2].axis("off")
cbar = fig.colorbar(im1, ax=axs[num_rows - 1, 2], orientation="vertical",
fraction=0.02, pad=0.04)
cbar.set_ticks(range(n_classes))
cbar.set_ticklabels(["background"] + list(template_all))
fig.savefig(save_path, bbox_inches="tight", dpi=200)
plt.close(fig)
def save_prob_maps(foreground_probs_all, template_all, save_path="prob_maps.npz"):
"""Save one probability map per class into a single .npz keyed by class name."""
assert len(foreground_probs_all) == len(template_all), \
"probability maps and class names must be the same length"
prob_dict = {name: prob.astype(np.float16)
for name, prob in zip(template_all, foreground_probs_all)}
np.savez_compressed(save_path, **prob_dict)
print(f"Saved probability maps to {save_path}")
def evaluate_segmentation(pred_mask, true_mask, eps=1e-7):
"""Binary segmentation metrics: (dice, accuracy, precision, recall)."""
pred_bin = (pred_mask > 0).astype(np.float32).reshape(-1)
true_bin = (true_mask > 0).astype(np.float32).reshape(-1)
intersection = np.sum(pred_bin * true_bin)
union = np.sum(pred_bin) + np.sum(true_bin)
dice = (2.0 * intersection + eps) / (union + eps)
acc = np.mean(pred_bin == true_bin)
tp = np.sum(pred_bin * true_bin)
fp = np.sum(pred_bin * (1.0 - true_bin))
fn = np.sum((1.0 - pred_bin) * true_bin)
precision = tp / (tp + fp + eps)
recall = tp / (tp + fn + eps)
return dice, acc, precision, recall
def compute_multi_class_metrics(gt, pred):
"""Macro-averaged IoU / Dice / precision / recall over the classes present in gt."""
metrics = {}
epsilon = 1e-7
num_classes = max(int(np.amax(gt)), int(np.amax(pred))) + 1
cm = confusion_matrix(gt.flatten(), pred.flatten(), labels=np.arange(num_classes))
IoU, Dice, Precision, Recall = [], [], [], []
for i in range(num_classes):
TP = cm[i, i]
FP = cm[:, i].sum() - TP
FN = cm[i, :].sum() - TP
# Skip classes not present in the GT
if (TP + FN) == 0:
continue
IoU.append(TP / (TP + FP + FN + epsilon))
Dice.append(2 * TP / (2 * TP + FP + FN + epsilon))
Precision.append(TP / (TP + FP + epsilon))
Recall.append(TP / (TP + FN + epsilon))
metrics["IoU"] = np.mean(IoU)
metrics["Dice"] = np.mean(Dice)
metrics["Precision"] = np.mean(Precision)
metrics["Recall"] = np.mean(Recall)
return metrics
|