hkust-aaron's picture
Add annotated sample pack: quality, timeline, clip captions, keyframe masks (v1/v2), SAM 3.1 dense tracks
07f5340 verified
Raw History Blame Contribute Delete
4.1 kB
"""Per-clip camera/scene motion statistics and offset overview sheets.
Usage: python 04_clip_motion.py <video> <out_csv> <sheet_dir>
The video is cut into 10 s clips (250 frames). Every 5th frame, corner features
are tracked to the frame 5 later (Lucas-Kanade) and a similarity transform
(shift + zoom + rotation) is fitted with RANSAC. Tissue deforms and instruments
move, so this is a rough whole-image motion estimate, not camera pose.
Per clip:
shift_px_per_s median image shift speed
inlier_ratio median share of tracks that agree with the fitted transform
motion_level low / medium / high = tercile of shift_px_per_s within this video
Only the relative ranking is meaningful. On this footage just 25-60 % of tracks
agree with a single transform (glare, deforming tissue, moving instruments), so
absolute speeds are inflated and a zoom/pan direction could not be estimated
reliably; an earlier zoom_in/zoom_out label was dropped for that reason.
Sheets: one thumbnail every 4 s starting at 2 s (the first overview set starts at 0 s).
"""
import csv
import sys
from pathlib import Path
from statistics import median
import cv2
import numpy as np
video, out_csv, sheet_dir = sys.argv[1], sys.argv[2], Path(sys.argv[3])
sheet_dir.mkdir(parents=True, exist_ok=True)
CLIP, STEP, FPS = 250, 5, 25
COLS, ROWS, TW, TH = 5, 6, 288, 230
cap = cv2.VideoCapture(video)
prev, idx, pairs, thumbs = None, 0, {}, []
while True:
ok, frame = cap.read()
if not ok:
break
if idx % 100 == 50:
t = cv2.resize(frame, (TW, TH), interpolation=cv2.INTER_AREA)
cv2.rectangle(t, (0, 0), (TW, 16), (0, 0, 0), -1)
cv2.putText(t, f"{idx / FPS:6.1f}s f{idx}", (4, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.4, (255, 255, 255), 1)
thumbs.append(t)
if idx % STEP == 0:
grey = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
if prev is not None:
p0 = cv2.goodFeaturesToTrack(prev, 300, 0.01, 12)
rec = None
if p0 is not None and len(p0) >= 20:
p1, st, _ = cv2.calcOpticalFlowPyrLK(prev, grey, p0, None, winSize=(21, 21), maxLevel=3)
good = st.reshape(-1) == 1
if good.sum() >= 20:
m, inl = cv2.estimateAffinePartial2D(p0[good], p1[good], method=cv2.RANSAC,
ransacReprojThreshold=3.0)
if m is not None:
scale = float(np.hypot(m[0, 0], m[1, 0]))
centre = np.array([360.0, 288.0])
shift = float(np.linalg.norm(m[:, :2] @ centre + m[:, 2] - centre))
rec = (shift, scale, float(inl.mean()))
pairs.setdefault((idx - STEP) // CLIP, []).append(rec)
prev = grey
idx += 1
cap.release()
with open(out_csv, "w", newline="") as f:
w = csv.writer(f)
w.writerow(["clip_id", "start_frame", "end_frame_exclusive", "shift_px_per_s", "inlier_ratio",
"pairs_fitted", "motion_level"])
shifts = sorted(median(r[0] for r in v if r) * FPS / STEP for v in pairs.values() if any(v))
t1, t2 = shifts[len(shifts) // 3], shifts[2 * len(shifts) // 3]
for c in sorted(pairs):
recs = [r for r in pairs[c] if r]
if not recs:
w.writerow([c, c * CLIP, min(idx, (c + 1) * CLIP), "", "", 0, "unknown"])
continue
shift = median(r[0] for r in recs) * FPS / STEP
inl = median(r[2] for r in recs)
level = "low" if shift < t1 else ("medium" if shift < t2 else "high")
w.writerow([c, c * CLIP, min(idx, (c + 1) * CLIP), round(shift, 1), round(inl, 2), len(recs), level])
per = COLS * ROWS
for s in range(0, len(thumbs), per):
sheet = np.zeros((ROWS * TH, COLS * TW, 3), np.uint8)
for i, t in enumerate(thumbs[s:s + per]):
y, x = divmod(i, COLS)
sheet[y * TH:(y + 1) * TH, x * TW:(x + 1) * TW] = t
cv2.imwrite(str(sheet_dir / f"offset_{s // per:02d}.jpg"), sheet, [cv2.IMWRITE_JPEG_QUALITY, 85])
print("frames", idx, "clips", len(pairs), "thumbs", len(thumbs))