5e2094f2ad
FastAPI service in infra/media-worker implementing every MediaInferenceProvider operation behind HTTP API v1, plus capabilities. Bearer auth, MEDIA_ROOT realpath enforcement, strict request schemas, cancellation on disconnect, GPU/CPU gates with 503 overflow, input size/duration limits, lazy SHA256-pinned model downloads, fail-closed redaction with atomic output, offline pytest suite with fake backends, CPU/CUDA/ROCm Dockerfile and a license review of every model. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0138YqDj3ttfayRwuMKKrqVj
59 lines
2.1 KiB
Python
59 lines
2.1 KiB
Python
"""saliency: spectral residual (Hou & Zhang, 2007), implemented with numpy/OpenCV.
|
|
|
|
The point is the saliency-weighted centroid of the top 5 % most salient
|
|
pixels; strength is how much those pixels stand out against the frame mean.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import threading
|
|
from typing import Any
|
|
|
|
import cv2
|
|
import numpy as np
|
|
|
|
from .. import media_io
|
|
from ..schemas import TimesRequest
|
|
from .common import Ctx, clamp01, open_input, unique_sorted
|
|
|
|
WORK_SIZE = 64
|
|
|
|
|
|
def saliency_map(frame: np.ndarray) -> np.ndarray:
|
|
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY).astype(np.float32)
|
|
small = cv2.resize(gray, (WORK_SIZE, WORK_SIZE), interpolation=cv2.INTER_AREA)
|
|
spectrum = np.fft.fft2(small)
|
|
log_amp = np.log(np.abs(spectrum) + 1e-8)
|
|
phase = np.angle(spectrum)
|
|
residual = log_amp - cv2.blur(log_amp, (3, 3))
|
|
sal = np.abs(np.fft.ifft2(np.exp(residual + 1j * phase))) ** 2
|
|
sal = cv2.GaussianBlur(sal.astype(np.float32), (0, 0), 2.5)
|
|
peak = float(sal.max())
|
|
return sal / peak if peak > 0 else sal
|
|
|
|
|
|
def salient_point(frame: np.ndarray) -> tuple[float, float, float]:
|
|
sal = saliency_map(frame)
|
|
threshold = float(np.quantile(sal, 0.95))
|
|
top = np.where(sal >= threshold, sal, 0.0)
|
|
total = float(top.sum())
|
|
if total <= 0:
|
|
return 0.5, 0.5, 0.0
|
|
ys, xs = np.mgrid[0 : sal.shape[0], 0 : sal.shape[1]]
|
|
cx = float((xs * top).sum() / total + 0.5) / sal.shape[1]
|
|
cy = float((ys * top).sum() / total + 0.5) / sal.shape[0]
|
|
mean = float(sal.mean())
|
|
top_mean = float(sal[sal >= threshold].mean())
|
|
strength = (top_mean - mean) / top_mean if top_mean > 0 else 0.0
|
|
return clamp01(cx), clamp01(cy), clamp01(strength)
|
|
|
|
|
|
def run(ctx: Ctx, body: TimesRequest, cancel: threading.Event) -> dict[str, Any]:
|
|
real, info = open_input(ctx, body.path, cancel)
|
|
ctx.registry.runtime("saliency")
|
|
points = []
|
|
for t_ms, frame in media_io.frames_at(real, info, unique_sorted(body.timesMs), cancel, 512):
|
|
cx, cy, strength = salient_point(frame)
|
|
points.append({"tMs": t_ms, "cx": cx, "cy": cy, "strength": strength})
|
|
return {"points": points}
|