"""Re-print the consultant's T-shirt logo on a tracked position — conservative erase. Plan JSON per frame: prints [[x,y,s,theta,width,weight],...], box [x0,x1,y0,y1] (template search area around the first print, in units of its scale), extra [[x0,x1,y0,y1],...] optional windows where stray print fragments are cleaned, sigma (print softness). Only three things are ever erased, so hair, skyline, collar and laptop are never touched: 1. old prints found by template matching the logo, and only their print-coloured pixels; 2. strictly magenta pixels of the chest marker (compact blobs); 3. print-coloured pixels in a tight window around the new print (and any `extra` window). Erased pixels are refilled from the white fabric around them, then the vector logo is printed once, behind hair and hands, tinted by the light on the shirt.""" import cv2, numpy as np, json, sys, os src, dst, plan = sys.argv[1], sys.argv[2], sys.argv[3] LOGO = os.environ.get('LOGO', 'logo_tee.png') # source/brand/logo_tee.png from the founder's package P = {int(k): v for k, v in json.load(open(plan)).items()} RAW = cv2.imread(LOGO, -1) logo = RAW.astype(np.float32); logo[..., :3] *= logo[..., 3:4] / 255.0 # premultiplied: no fringe after warp/blur lh, lw = logo.shape[:2] a_ = RAW[..., 3:4] / 255.0 LG = cv2.cvtColor((RAW[..., :3] * a_ + 255 * (1 - a_)).astype(np.uint8), cv2.COLOR_BGR2GRAY) # the logo printed on white def quad(x, y, th, pts): c, s_ = np.cos(th), np.sin(th) return np.array([[x + c * px - s_ * py, y + s_ * px + c * py] for px, py in pts], np.int32) def rect(H, W, x, y, th, s, r): m = np.zeros((H, W), np.uint8); x0, x1, y0, y1 = [v * s for v in r] cv2.fillPoly(m, [quad(x, y, th, [(x0, y0), (x1, y0), (x1, y1), (x0, y1)])], 1); return m def on_shirt(bb, hsv0, b, r, base): """True when the band around a box is white fabric (not blue sky, not a dark scene).""" ring = (cv2.dilate(bb, np.ones((11, 11), np.uint8)) - bb) > 0 return np.median(hsv0[..., 2][ring]) > 185 and np.median(hsv0[..., 1][ring]) < 50 and np.median((b - r)[ring]) < base + 8 stats = {} for i in sorted(P): e = P[i]; fr = cv2.imread(f'{src}/f{i:03d}.png'); H, W = fr.shape[:2] x, y, s, th, LW = e['prints'][0][:5] ox, oy = e.get('offset', [0, 0]) prints = [] for (px, py, ps, pth, LWp, wgt) in e['prints']: prints.append([px + ps * (ox * np.cos(pth) - oy * np.sin(pth)), py + ps * (ox * np.sin(pth) + oy * np.cos(pth)), ps, pth, e.get('width', LWp), wgt]) if 'print_at' in e: # pinned to her neckline (body tracking drifts when she turns) prints = [[e['print_at'][0], e['print_at'][1], 1.0, 0.0, e['print_w'], 1.0]] # one print, absolute position and width in px nx, ny, _, _, NW = prints[0][:5] f = fr.astype(np.float32); b, g, r = f[..., 0], f[..., 1], f[..., 2]; V = f.max(2) reg = rect(H, W, x, y, th, s, e['box']) br = (reg > 0) & (V > 200) base = float(np.median((b - r)[br])) if br.sum() > 50 else 4.0 vmed = float(np.median(V[br])) if br.sum() > 50 else 240 hair = (r - b > 10) & (g > b - 25) # hair and skin are warm; the print never is blk = (V < 100) & (r - b < 15) # dark, not brown n, lab, st, _ = cv2.connectedComponentsWithStats(cv2.erode(blk.astype(np.uint8), np.ones((5, 5), np.uint8))) big = np.isin(lab, [k for k in range(1, n) if st[k][4] > 300 * s * s or max(st[k][2], st[k][3]) > 160 * s]) bezel = (cv2.dilate(big.astype(np.uint8), np.ones((7, 7), np.uint8)) > 0) & blk # a long dark shape (laptop bezel); letters are small laptops = [st[k][:4] for k in range(1, n) if max(st[k][2], st[k][3]) > 160 * s] # bounding boxes of long dark frames: text inside is a screen printcol = (((b - r - base > 14) & (b > g - 10)) | ((V < vmed - 70) & (b - r > 8))) & ~hair & ~bezel fg = np.zeros((H, W), np.uint8) hsv0 = cv2.cvtColor(fr, cv2.COLOR_BGR2HSV) # 'soft' erase (greyed letter cores) only well inside the shirt: an arm or sleeve hem in cool light is dark and bluish too inshirt = cv2.GaussianBlur(((V > 165) & (hsv0[..., 1] < 60)).astype(np.float32), (0, 0), 7 * max(s, .6)) > 0.6 # 1. whole old prints, by template matching at any size; erase their print-coloured (and greyed-core) pixels ys, xs = np.where(reg > 0); rx0, ry0, rx1, ry1 = xs.min(), ys.min(), xs.max() + 1, ys.max() + 1 sub = cv2.cvtColor(fr, cv2.COLOR_BGR2GRAY)[ry0:ry1, rx0:rx1].copy(); nm = 0 for _ in range(3): best = (0, None) for wpx in np.geomspace(36, 260, 34): hpx = max(6, int(round(wpx * lh / lw))); wpx = int(round(wpx)) if hpx >= sub.shape[0] or wpx >= sub.shape[1]: continue res = cv2.matchTemplate(sub, cv2.resize(LG, (wpx, hpx), interpolation=cv2.INTER_AREA), cv2.TM_CCOEFF_NORMED) _, mv, _, ml = cv2.minMaxLoc(res) if mv > best[0]: best = (mv, (ml[0], ml[1], wpx, hpx)) if best[0] < 0.55: break bx, by, bw, bh = best[1]; box = np.zeros_like(reg) cv2.rectangle(box, (rx0 + bx - 2, ry0 + by - 2), (rx0 + bx + bw + 2, ry0 + by + bh + 2), 1, -1) if on_shirt(box, hsv0, b, r, base): # a 'match' in the skyline is not a print fg |= ((box > 0) & (printcol | ((V < vmed - 25) & (b - r > 2) & ~hair & ~bezel & inshirt))).astype(np.uint8); nm += 1 sub[max(0, by - 2):by + bh + 2, max(0, bx - 2):bx + bw + 2] = int(np.median(sub)) # 1b. old prints the template misses (each was warped a little differently): a word-shaped cluster of print-coloured # pixels sitting on white fabric. Hair is warm and the skyline is not ringed by white shirt, so neither qualifies. wreg = reg if 'word' not in e else np.max([rect(H, W, x, y, th, s, rr) for rr in e['word']], axis=0) pc = ((printcol | ((V < vmed - 45) & (b - r > 2) & ~hair & ~bezel)) & (wreg > 0)).astype(np.uint8) word = cv2.morphologyEx(pc, cv2.MORPH_CLOSE, cv2.getStructuringElement(cv2.MORPH_RECT, (int(13 * max(s, .5)) | 1, 3))) n, lab, st, _ = cv2.connectedComponentsWithStats(word) for k in range(1, n): bx_, by_, bw_, bh_, ar_ = st[k] if not (30 * s <= bw_ <= 220 * s and 5 <= bh_ <= 45 * s and bw_ >= 2.5 * bh_ and ar_ >= 0.12 * bw_ * bh_): continue cxw, cyw = bx_ + bw_ / 2, by_ + bh_ / 2 if any(lx <= cxw <= lx + lw_ and ly <= cyw <= ly + lh_ for lx, ly, lw_, lh_ in laptops): continue # laptop screen text bb = np.zeros((H, W), np.uint8); cv2.rectangle(bb, (bx_ - 3, by_ - 3), (bx_ + bw_ + 3, by_ + bh_ + 3), 1, -1) if on_shirt(bb, hsv0, b, r, base): fg |= ((bb > 0) & (printcol | ((V < vmed - 25) & (b - r > 2) & ~hair & ~bezel & inshirt))).astype(np.uint8); nm += 1 # 2. the magenta chest marker: strictly magenta, compact, and inside the box (a ribbon always crosses its edge); # then its soft fringe (pink on white, lilac under the cyan sweep, dark magenta where it meets hair) core = (r - g > 30) & (b - g > 28) if e.get('marker_soft'): core |= (r - g > 10) & (b - g > 8) & (np.abs(r - b) < 18) # washed-out marker in a light sweep mg = cv2.dilate((core & (reg > 0)).astype(np.uint8), np.ones((3, 3), np.uint8)) edge = cv2.dilate(reg, np.ones((3, 3), np.uint8)) - cv2.erode(reg, np.ones((3, 3), np.uint8)) n, lab, st, _ = cv2.connectedComponentsWithStats(mg); mk = np.zeros((H, W), np.uint8) for k in range(1, n): cm = (lab == k).astype(np.uint8) if st[k][4] >= 12 and st[k][2] < 160 * s and st[k][3] < 90 * s and not (cm & edge).any(): mk |= cm if mk.any(): near = cv2.dilate(mk, np.ones((int(27 * max(s, .6)) | 1,) * 2, np.uint8)) > 0 tinge = ((r - g > 6) & (b - g > 4) & (r - b < 20)) | ((b - g > 10) & (r - g > 8)) # pink / purple if e.get('marker_soft'): tinge |= (r - g > 2) & (b - g > 0) # lilac on a cyan-washed shirt (shirt itself has r < g) fg |= (mk > 0).astype(np.uint8) fg |= (near & tinge & (hsv0[..., 1] < 90)).astype(np.uint8) # 3. stray print fragments right where the new print goes, and in any extra window lhh = NW * lh / lw / 2 ws = prints[0][2] win = rect(H, W, nx, ny, 0 if 'print_at' in e else th, ws, [-(NW / 2 + 30 * s / ws), NW / 2 + 30 * s / ws, -(lhh + 14 * s / ws), lhh + 14 * s / ws]) for rr in e.get('extra', []): win |= rect(H, W, x, y, th, s, rr) fg |= ((printcol | ((V < vmed - 40) & (b - r > 2) & ~hair & ~bezel)) & (win > 0)).astype(np.uint8) # incl. greyed letter cores fg = cv2.dilate(fg, np.ones((3, 3), np.uint8)); fg[bezel] = 0 # 'force' windows: an old print lying ON the laptop's edge — navy/blue ink there is erased even next to the bezel for rr in e.get('force', []): fw = rect(H, W, x, y, th, s, rr) > 0 ink_ = ((b - r - base > 14) | ((V < vmed - 70) & (b - r > 8))) & ~hair & ~((V < 60) & (np.abs(b - r) < 6)) fg |= cv2.dilate((fw & ink_).astype(np.uint8), np.ones((3, 3), np.uint8)) # absolute-pixel windows for stray fragments: dark ink, royal-blue ink (g above r; the purple ribbon has r above g), hot-magenta marker for (ax0, ax1, ay0, ay1) in e.get('abs', []): aw = np.zeros((H, W), bool); aw[ay0:ay1, ax0:ax1] = True ink_ = ((V < vmed - 25) & ~hair & (b > r - 5)) | ((b - r > 20) & (g - r > 5)) | ((r > 180) & (g < 150) & (b > 160) & (r - g > 60)) # incl. anti-aliased letter edges fg |= cv2.dilate((aw & ink_ & ~bezel).astype(np.uint8), np.ones((5, 5), np.uint8)) fg[hair & (V < 235)] = 0 # never paint over skin or hair (sleeve hems, hair edges) for (ax0, ax1, ay0, ay1) in e.get('abs_dark', []): # an old print lying over bare skin (sleeve hem): anything dark in a small window aw = np.zeros((H, W), bool); aw[ay0:ay1, ax0:ax1] = True fg |= cv2.dilate((aw & (V < 150)).astype(np.uint8), np.ones((5, 5), np.uint8)) if fg.any(): fab = ((fg == 0) & (V > 165) & (cv2.cvtColor(fr, cv2.COLOR_BGR2HSV)[..., 1] < 60)).astype(np.float32); sg = 5 * max(s, .5) num = cv2.GaussianBlur(f * fab[..., None], (0, 0), sg); den = cv2.GaussianBlur(fab, (0, 0), sg)[..., None] fill = num / np.maximum(den, 1e-4); ok = (den[..., 0] > .45) & (fg > 0) # mostly fabric around: fabric fill; else (hair, skin) plain inpaint out = cv2.inpaint(fr, fg, 4, cv2.INPAINT_TELEA).astype(np.float32) nc, lc, sc_, _ = cv2.connectedComponentsWithStats(fg) # fabric fill only for blobs ringed by fabric; else plain inpaint for k in range(1, nc): cm = (lc == k).astype(np.uint8); ring = (cv2.dilate(cm, np.ones((9, 9), np.uint8)) - cm) > 0 if fab[ring].mean() > 0.85: m_ = (cm > 0) & ok; out[m_] = fill[m_] fr = np.clip(out, 0, 255).astype(np.uint8) f = fr.astype(np.float32) hsv = cv2.cvtColor(fr, cv2.COLOR_BGR2HSV).astype(np.float32); Vv, Sa = hsv[..., 2], hsv[..., 1] shirt = np.clip((Vv - 130) / 50, 0, 1) * np.clip((70 - Sa) / 35, 0, 1); shirt = cv2.GaussianBlur(shirt, (0, 0), 1.0) light = cv2.GaussianBlur(f, (0, 0), 6) tint = np.clip(light / np.maximum(light.max(2, keepdims=True), 1), 0, 1) * np.clip(cv2.GaussianBlur(Vv, (0, 0), 6) / 238.0, 0.55, 1.0)[..., None] wash = np.clip(cv2.GaussianBlur(Sa, (0, 0), 6) / 110.0, 0, 0.35)[..., None] # a laptop held in front of her hides the shirt: everything below a bezel's top edge, across its width if laptops: cover = np.zeros((H, W), np.uint8) nb, lb, sb, _ = cv2.connectedComponentsWithStats(bezel.astype(np.uint8)) for k in range(1, nb): if sb[k][2] < 120 * s or sb[k][2] < 1.5 * sb[k][3] or sb[k][1] < ny - 20 * s: continue # a laptop edge: wide, flat, below the print (not hair) ys_, xs_ = np.where(lb == k) for cx_ in range(xs_.min(), xs_.max() + 1): col = ys_[xs_ == cx_] if col.size: cover[col.min():, cx_] = 1 shirt = shirt * (1 - cv2.GaussianBlur(cover.astype(np.float32), (0, 0), 1.0)) for (px, py, ps, pth, LWp, wgt) in prints: if wgt <= 0: continue M = cv2.getRotationMatrix2D((lw / 2, lh / 2), -np.degrees(pth), LWp * ps / lw); M[:, 2] += [px - lw / 2, py - lh / 2] wl = cv2.GaussianBlur(cv2.warpAffine(logo, M, (W, H), flags=cv2.INTER_AREA, borderValue=(0, 0, 0, 0)), (0, 0), e['sigma']) a = (wl[..., 3] / 255.0) * shirt * 0.95 * wgt rgb = np.clip(wl[..., :3] / np.maximum(wl[..., 3:4] / 255.0, 1e-3), 0, 255) * tint rgb = rgb * (1 - wash) + light * wash f = f * (1 - a[..., None]) + rgb * a[..., None] cv2.imwrite(f'{dst}/f{i:03d}.png', np.clip(f, 0, 255).astype(np.uint8)) stats[i] = (nm, int(fg.sum())) print('done', len(P), 'frames; template hits per frame:', ''.join(str(v[0]) for k, v in sorted(stats.items())))