def resolve_image_path(name):
candidates = [Path(name), Path("posts") / name]
for path in candidates:
if path.exists():
return path
raise FileNotFoundError(name)
def load_image_local(name):
return Image.open(resolve_image_path(name)).convert("RGB")
def xyxy_from_center_size(x, y, w, h, image_w, image_h):
cx, cy = x * image_w, y * image_h
bw, bh = w * image_w, h * image_h
return [cx - bw / 2, cy - bh / 2, cx + bw / 2, cy + bh / 2]
def mask_to_measurements(mask):
mask = np.asarray(mask).astype(bool)
ys, xs = np.nonzero(mask)
if len(xs) == 0:
raise ValueError("Empty mask")
return {
"xyxy": [float(xs.min()), float(ys.min()), float(xs.max()), float(ys.max())],
"centroid": [float(xs.mean()), float(ys.mean())],
"area_px": int(mask.sum()),
}
def falcon_detection_to_object(det, label, image_size):
width, height = image_size
if "mask" in det and det["mask"] is not None:
base = mask_to_measurements(det["mask"])
else:
xyxy = xyxy_from_center_size(
det["xy"]["x"], det["xy"]["y"], det["hw"]["w"], det["hw"]["h"], width, height
)
base = {
"xyxy": xyxy,
"centroid": [float((xyxy[0] + xyxy[2]) / 2), float((xyxy[1] + xyxy[3]) / 2)],
"area_px": int(max(0.0, xyxy[2] - xyxy[0]) * max(0.0, xyxy[3] - xyxy[1])),
}
return {**base, "label": label}
def choose_primary(objects, label):
matches = [obj for obj in objects if obj["label"] == label]
if not matches:
raise ValueError(f"No objects found for {label!r}")
return max(matches, key=lambda obj: obj["area_px"])
def objects_to_detections(objects):
if not objects:
return sv.Detections(
xyxy=np.zeros((0, 4), dtype=np.float32),
class_id=np.array([], dtype=int),
data={"class_name": []},
)
label_to_id = {label: idx for idx, label in enumerate(sorted({obj["label"] for obj in objects}))}
xyxy = np.array([obj["xyxy"] for obj in objects], dtype=np.float32)
class_id = np.array([label_to_id[obj["label"]] for obj in objects], dtype=int)
class_name = [obj["label"] for obj in objects]
return sv.Detections(xyxy=xyxy, class_id=class_id, data={"class_name": class_name})
BOX_ANNOTATOR = sv.BoxAnnotator(thickness=2, color_lookup=sv.ColorLookup.CLASS)
LABEL_ANNOTATOR = sv.LabelAnnotator(
text_scale=0.45,
text_padding=4,
color_lookup=sv.ColorLookup.CLASS,
)
def render_annotated_scene(image, objects, show_labels=True):
scene = np.array(image.convert("RGB")).copy()
if not objects:
return scene
detections = objects_to_detections(objects)
scene = BOX_ANNOTATOR.annotate(scene=scene, detections=detections)
if show_labels:
labels = [obj["label"] for obj in objects]
scene = LABEL_ANNOTATOR.annotate(scene=scene, detections=detections, labels=labels)
return scene
def _overlay_text(ax, x, y, text, ha="left"):
ax.text(
x,
y,
text,
ha=ha,
va="top",
color="white",
fontsize=11,
bbox={"facecolor": "black", "alpha": 0.65, "pad": 4},
)
def plot_result_scene(result):
measurement = result["measurement"]
analysis = measurement["analysis"]
focus_objects = result["focus_objects"]
show_labels = len(focus_objects) <= 8
scene = render_annotated_scene(result["image"], focus_objects, show_labels=show_labels)
fig, ax = plt.subplots(figsize=(10, 6))
ax.imshow(scene)
if analysis == "count_by_half":
split_x = measurement["split_x"]
ax.axvline(split_x, color="#00d4ff", linestyle="--", linewidth=2)
_overlay_text(ax, split_x - 14, 26, f"left: {measurement['left_count']}", ha="right")
_overlay_text(ax, split_x + 14, 26, f"right: {measurement['right_count']}")
elif analysis == "width_threshold":
_overlay_text(
ax,
18,
26,
f"width fraction = {measurement['width_fraction']:.4f} vs threshold {measurement['threshold']:.4f}",
)
elif analysis == "edge_margin_threshold":
obj = focus_objects[0]
x1, y1, x2, y2 = obj["xyxy"]
edge = measurement["edge"]
y_mid = (y1 + y2) / 2
if edge == "right":
ax.hlines(y_mid, x2, result["image"].size[0], colors="#ff4d6d", linewidth=3)
_overlay_text(
ax,
result["image"].size[0] - 18,
y_mid - 24,
f"right margin {measurement['margin_px']} px vs threshold {measurement['threshold_px']} px",
ha="right",
)
elif edge == "left":
ax.hlines(y_mid, 0, x1, colors="#00d4ff", linewidth=3)
_overlay_text(ax, 18, y_mid - 24, f"left margin {measurement['margin_px']} px vs threshold {measurement['threshold_px']} px")
elif analysis == "centered_square_crop":
crop_x1, crop_y1, crop_x2, crop_y2 = measurement["square_crop_xyxy"]
rect = patches.Rectangle(
(crop_x1, crop_y1),
crop_x2 - crop_x1,
crop_y2 - crop_y1,
linewidth=3,
edgecolor="#00d4ff",
facecolor="none",
linestyle="--",
)
ax.add_patch(rect)
_overlay_text(ax, crop_x1 + 18, crop_y1 + 34, "centered square crop")
ax.set_title(result["task"]["name"], fontsize=13)
ax.axis("off")
plt.tight_layout()
plt.show()
def show_task_gallery(tasks):
fig, axes = plt.subplots(2, 2, figsize=(12, 8))
for ax, task in zip(axes.flat, tasks):
ax.imshow(load_image_local(task["image"]))
ax.set_title(task["name"], fontsize=11)
ax.set_xlabel(task["question"], fontsize=9)
ax.axis("off")
plt.tight_layout()
plt.show()
def show_task_gallery_extended(tasks):
"""Show task gallery with question overlaid on each image."""
import textwrap
n = len(tasks)
cols = min(n, 4)
rows = (n + cols - 1) // cols
fig, axes = plt.subplots(rows, cols, figsize=(15, 5 * rows))
if rows == 1 and cols == 1:
axes = [axes]
elif rows == 1:
axes = list(axes)
else:
axes = list(axes.flat)
for i, task in enumerate(tasks):
img = load_image_local(task['image'])
axes[i].imshow(img)
# Overlay question at the top of the image
q = textwrap.fill(task['question'], width=35)
axes[i].text(
0.5, 0.97, q,
transform=axes[i].transAxes, fontsize=9, fontweight='bold',
color='white', ha='center', va='top',
bbox=dict(boxstyle='round,pad=0.4', facecolor='black', alpha=0.75),
)
# Overlay expected answer at bottom
exp = task.get('expected')
if exp is not None:
exp_str = 'yes' if exp is True else 'no' if exp is False else str(exp)
axes[i].text(
0.5, 0.03, f'Expected: {exp_str}',
transform=axes[i].transAxes, fontsize=8,
color='white', ha='center', va='bottom',
bbox=dict(boxstyle='round,pad=0.3', facecolor='#166534', alpha=0.8),
)
axes[i].set_title(task['name'], fontsize=11, fontweight='bold', pad=8)
axes[i].axis('off')
for j in range(n, len(axes)):
axes[j].axis('off')
plt.tight_layout()
plt.show()
