tools/dump_frames.py

"""
Decode selected frame ranges of the test video, downscale to the detector's
analysis resolution, and write raw RGBA frames + timing so the ACTUAL
detector.js can be run in Node against real data.

Binary format:
  header: <iiii  aw, ah, fps_int, count
  then per frame: <i frameIndex, then aw*ah*4 bytes RGBA
"""
import cv2, numpy as np, struct

PATH = r"E:\test\Kiro_flash_detection\Video\0921 (1)_flash.mp4"
OUT = r"E:\test\Kiro_flash_detection\tools\frames.bin"
ANALYSIS_W = 480

# Include >=15 lead-in frames before each window for the rolling background.
WINDOWS = [
    (2580, 2660),   # the real flash at 2612 (must still be detected)
    (2330, 2380),   # the reported false positive at 2353-2358 / peak 2355
    (2510, 2560),   # prior false-positive cluster ~2535
    (2080, 2160),   # prior false-positive cluster ~2105
]

cap = cv2.VideoCapture(PATH)
fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
W = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)); H = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
scale = W/ANALYSIS_W if W > ANALYSIS_W else 1.0
aw = max(1, round(W/scale)); ah = max(1, round(H/scale))

want, seen = [], set()
for a, b in WINDOWS:
    for i in range(a, b+1):
        if i not in seen:
            seen.add(i); want.append(i)
want.sort()

records = []
for idx in want:
    cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
    ok, f = cap.read()
    if not ok: continue
    s = cv2.resize(f, (aw, ah), interpolation=cv2.INTER_AREA)
    rgba = cv2.cvtColor(s, cv2.COLOR_BGR2RGBA)
    records.append((idx, rgba.tobytes()))
cap.release()

with open(OUT, "wb") as fo:
    fo.write(struct.pack("<iiii", aw, ah, int(round(fps)), len(records)))
    for idx, buf in records:
        fo.write(struct.pack("<i", idx))
        fo.write(buf)

print(f"wrote {len(records)} frames {aw}x{ah} fps~{fps} scale={scale:.3f} to {OUT}")