#!/usr/bin/env python3 """Measure the rendered caption frames produced by test-run.sh (stdlib only). Reads the raw rgb24 frames that ffmpeg wrote and reports: A. Cap height of the sample caption at each font size, on a 1080x1920 frame and after the frame is scaled to a 390 CSS px wide phone view. B. The background under the caption on two real clips: how bright it gets, and what contrast a caption of each style has against the colour its letters actually touch. C. How often the caption band of a real clip contains a pixel as bright as white text (from the ffmpeg signalstats log). """ import os import re import sys OUT = sys.argv[1] FONT = sys.argv[2] PHONE_W = int(sys.argv[3]) FRAME_W = int(sys.argv[4]) FRAME_H = int(sys.argv[5]) TEXT = sys.argv[6] SCALE = PHONE_W / FRAME_W ROW = FRAME_W * 3 STYLES = [ ("white-plain", (255, 255, 255)), ("white-shadow", (255, 255, 255)), ("white-outline", (255, 255, 255)), ("white-box60", (255, 255, 255)), ("white-box80", (255, 255, 255)), ("white-box100", (255, 255, 255)), ("yellow-box60", (255, 226, 74)), ] def load(name): with open(os.path.join(OUT, name), "rb") as fh: data = fh.read() assert len(data) == ROW * FRAME_H, "%s: expected %d bytes, got %d" % (name, ROW * FRAME_H, len(data)) return data def lum(r, g, b): def ch(c): cs = c / 255.0 return cs / 12.92 if cs <= 0.04045 else ((cs + 0.055) / 1.055) ** 2.4 return 0.2126 * ch(r) + 0.7152 * ch(g) + 0.0722 * ch(b) def ratio(a, b): la, lb = lum(*a), lum(*b) hi, lo = (la, lb) if la >= lb else (lb, la) return (hi + 0.05) / (lo + 0.05) def pct(vals, p): if not vals: return 0.0 v = sorted(vals) return v[max(0, min(len(v) - 1, int(round((p / 100.0) * (len(v) - 1)))))] def close(c, fill, tol): return abs(c[0] - fill[0]) <= tol and abs(c[1] - fill[1]) <= tol and abs(c[2] - fill[2]) <= tol print("=" * 100) print("A. CAP HEIGHT BY FONT SIZE sample caption: %r" % TEXT) print("=" * 100) print("font file: %s" % FONT) print("frame: %dx%d, phone view: %d CSS px wide (scale %.5f)" % (FRAME_W, FRAME_H, PHONE_W, SCALE)) print("%-9s %-14s %-16s %-16s %-17s %-16s %-14s" % ("size_px", "cap_height_px", "text_width_px", "pct_frame_width", "cap_css_390px", "pct_of_844px", "chars_fit")) size_rows = [] for name in sorted(os.listdir(OUT)): if not (name.startswith("size-") and name.endswith(".raw")): continue size = int(name[5:-4]) data = load(name) top = bottom = None for y in range(FRAME_H): row = data[y * ROW:(y + 1) * ROW] if row.count(0) != len(row): if top is None: top = y bottom = y cap = bottom - top + 1 # width of the same caption at this size left = right = None for x in range(FRAME_W): hit = False for y in range(top, bottom + 1): o = y * ROW + x * 3 if data[o] or data[o + 1] or data[o + 2]: hit = True break if hit: if left is None: left = x right = x width = right - left + 1 n_chars = len(TEXT) # characters that fit in 80% of the frame width at this size per_char = width / float(n_chars) chars_fit = int((FRAME_W * 0.8) // per_char) row = {"size_px": size, "cap_height_px": cap, "cap_ratio": round(cap / size, 4), "pct_frame_height": round(100.0 * cap / FRAME_H, 2), "text_width_px": width, "pct_frame_width": round(100.0 * width / FRAME_W, 1), "chars_fit_in_80pct": chars_fit, "cap_css_390px": round(cap * SCALE, 2)} size_rows.append(row) print("%-9d %-14d %-16d %-16s %-17s %-16s %-14s" % ( row["size_px"], row["cap_height_px"], row["text_width_px"], str(row["pct_frame_width"]) + "%", str(row["cap_css_390px"]) + " px", str(round(100.0 * row["cap_css_390px"] / 844, 2)) + "%", str(chars_fit) + " of " + str(n_chars))) print() print("=" * 100) print("B. WHAT THE CAPTION LETTERS ACTUALLY TOUCH, ON TWO REAL CLIPS") print("=" * 100) print("A support pixel is solid when its four neighbours have almost the same brightness") print("(that rejects antialiased letter edges). 'Worst' is the lowest contrast found") print("between the caption fill and a support pixel next to it.") print() contrast_rows = [] for clip in ("dark", "bright"): plain = load("%s-plain.raw" % clip) # background brightness in the caption area, with the caption removed bg_lums = [] for y in range(1180, 1440): base = y * ROW for x in range(0, FRAME_W): o = base + x * 3 bg_lums.append(lum(plain[o], plain[o + 1], plain[o + 2])) bg_max = max(bg_lums) bg_p90 = pct(bg_lums, 90) bg_med = pct(bg_lums, 50) print("-" * 100) print("%s clip: background in the caption band, caption removed: brightest pixel %.3f, " "90th percentile %.3f, median %.3f" % (clip, bg_max, bg_p90, bg_med)) bg_max_rgb = (0, 0, 0) bg_max_val = -1.0 for y in range(1180, 1440): base = y * ROW for x in range(FRAME_W): o = base + x * 3 c = (plain[o], plain[o + 1], plain[o + 2]) l = lum(*c) if l > bg_max_val: bg_max_val = l bg_max_rgb = c print(" a plain white caption has no separation there: %.2f:1 against the brightest pixel " "in the band (RGB %d,%d,%d)" % (ratio((255, 255, 255), bg_max_rgb), bg_max_rgb[0], bg_max_rgb[1], bg_max_rgb[2])) for name, fill in STYLES: path = "%s-%s.raw" % (clip, name) if not os.path.exists(os.path.join(OUT, path)): continue styled = load(path) # solid fill pixels fill_pixels = [] for y in range(1180, 1440): base = y * ROW for x in range(FRAME_W): o = base + x * 3 c = (styled[o], styled[o + 1], styled[o + 2]) if close(c, fill, 30): fill_pixels.append((x, y)) if not fill_pixels: continue ratios = [] for (x, y) in fill_pixels: best = None for dx, dy in ((1, 0), (-1, 0), (0, 1), (0, -1), (2, 0), (-2, 0), (0, 2), (0, -2)): nx, ny = x + dx, y + dy if not (0 <= nx < FRAME_W and 0 <= ny < FRAME_H): continue o = ny * ROW + nx * 3 c = (styled[o], styled[o + 1], styled[o + 2]) if close(c, fill, 30): continue # solid test: the four neighbours of this candidate must be close in brightness neigh = [] for ex, ey in ((1, 0), (-1, 0), (0, 1), (0, -1)): qx, qy = nx + ex, ny + ey if not (0 <= qx < FRAME_W and 0 <= qy < FRAME_H): neigh = None break q = qy * ROW + qx * 3 neigh.append(lum(styled[q], styled[q + 1], styled[q + 2])) if not neigh: continue if max(neigh) - min(neigh) > 0.08: continue # antialiased edge, not a solid support colour if best is None or ratio(fill, c) < ratio(fill, best): best = c if best is not None: ratios.append(ratio(fill, best)) row = {"clip": clip, "style": name, "fill": "#%02X%02X%02X" % fill, "fill_pixels": len(fill_pixels), "contrast_median": round(pct(ratios, 50), 2), "contrast_p10": round(pct(ratios, 10), 2), "contrast_min": round(min(ratios), 2), "bg_max_lum": round(bg_max, 3)} contrast_rows.append(row) print(" %-15s fill %s contrast against what the letters touch: median %6.2f:1, " "worst 10%% %6.2f:1, lowest %6.2f:1" % ( name, row["fill"], row["contrast_median"], row["contrast_p10"], row["contrast_min"])) print() print("=" * 100) print("C. HOW BRIGHT IS THE CAPTION BAND OVER TIME? (ffmpeg signalstats)") print("=" * 100) band_rows = [] for clip in ("dark", "bright"): path = os.path.join(OUT, "stats-%s.txt" % clip) if not os.path.exists(path): continue ymax, yavg = [], [] for line in open(path): m = re.search(r"lavfi\.signalstats\.YMAX=([0-9.]+)", line) if m: ymax.append(float(m.group(1))) m = re.search(r"lavfi\.signalstats\.YAVG=([0-9.]+)", line) if m: yavg.append(float(m.group(1))) if not ymax: continue bright_frames = sum(1 for v in ymax if v >= 250) row = {"clip": clip, "frames": len(ymax), "frames_with_white_pixel": bright_frames, "pct": round(100.0 * bright_frames / max(1, len(ymax)), 1), "ymax_median": round(pct(ymax, 50), 1), "ymax_max": round(max(ymax), 1), "yavg_median": round(pct(yavg, 50), 1)} band_rows.append(row) print("%-7s %3d frames at 10 fps. %3d frames (%.1f%%) contain a pixel as bright as white text " "(YMAX 255). Median YMAX %.0f, max %.0f, median band brightness YAVG %.0f." % ( clip, row["frames"], row["frames_with_white_pixel"], row["pct"], row["ymax_median"], row["ymax_max"], row["yavg_median"])) print() print("=" * 100) print("D. SUMMARY") print("=" * 100) print("cap height rows: %d, contrast rows: %d, band rows: %d" % ( len(size_rows), len(contrast_rows), len(band_rows)))