3467dbcf03
Full-stack application for predicting where humans look in images using DeepGaze saliency models. Includes heatmap overlays, gaze sequence prediction, hotspot detection, AOI analysis, rule-based insights, optional Claude AI design analysis, and professional PDF report generation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
42 lines
1.2 KiB
Python
42 lines
1.2 KiB
Python
import numpy as np
|
|
from scipy.ndimage import gaussian_filter
|
|
|
|
|
|
def extract_gaze_sequence(
|
|
saliency: np.ndarray,
|
|
num_fixations: int = 5,
|
|
inhibition_radius_fraction: float = 0.1,
|
|
) -> list[dict]:
|
|
"""
|
|
Extract predicted fixation sequence using iterative peak finding
|
|
with inhibition-of-return.
|
|
"""
|
|
sal = saliency.copy().astype(np.float64)
|
|
h, w = sal.shape
|
|
inhibition_radius = int(max(h, w) * inhibition_radius_fraction)
|
|
fixations = []
|
|
|
|
for rank in range(1, num_fixations + 1):
|
|
smoothed = gaussian_filter(sal, sigma=max(h, w) * 0.01)
|
|
if smoothed.max() < 1e-10:
|
|
break
|
|
|
|
peak_idx = np.unravel_index(np.argmax(smoothed), smoothed.shape)
|
|
y, x = int(peak_idx[0]), int(peak_idx[1])
|
|
prob = float(saliency[y, x])
|
|
|
|
fixations.append({
|
|
"rank": rank,
|
|
"x": x,
|
|
"y": y,
|
|
"x_pct": round(x / w * 100, 1),
|
|
"y_pct": round(y / h * 100, 1),
|
|
"probability": round(prob, 4),
|
|
})
|
|
|
|
# Inhibition of return
|
|
yy, xx = np.ogrid[:h, :w]
|
|
mask = (xx - x) ** 2 + (yy - y) ** 2 <= inhibition_radius ** 2
|
|
sal[mask] = 0.0
|
|
|
|
return fixations
|