"""Draw RAQEEB detections as labelled boxes (class name + confidence) and make before/after images. python annotate_raqeeb.py plan.png other.pdf --weights raqeeb-cctv-v1.0.0.pt --thresholds thresholds.json --out out_dir For every page it writes [_p]_annotated.jpg (boxes and labels), [_p]_before_after.jpg (original | annotated) and [_p]_boxes.json (box corners in image pixels, class, confidence). Detection is identical to infer_raqeeb.py (same thresholds, whole page by default, optional --tile). """ import argparse import json from pathlib import Path from types import SimpleNamespace from PIL import Image, ImageDraw, ImageFont NAMES = ['wall', 'door', 'window', 'open_passage', 'column', 'stair', 'elevator', 'escalator', 'exit_marker', 'gate', 'space_region'] COLORS = [(235, 60, 60), (20, 170, 60), (40, 100, 240), (255, 140, 30), (130, 60, 220), (0, 160, 170), (190, 60, 140), (170, 140, 0), (240, 30, 190), (80, 120, 40), (110, 110, 210)] def _font(size): for name in ('arial.ttf', 'DejaVuSans.ttf'): try: return ImageFont.truetype(name, size) except OSError: pass try: from matplotlib import font_manager return ImageFont.truetype(font_manager.findfont('DejaVu Sans'), size) except Exception: return ImageFont.load_default() def boxes_from_detections(detections): """infer_raqeeb detections (shapely geometry in raster px) -> list of dicts with xyxy boxes.""" out = [] for d in detections: x0, y0, x1, y1 = d['geometry'].bounds out.append(dict(class_id=int(d['class_id']), label=NAMES[int(d['class_id'])], confidence=round(float(d['confidence']), 4), box_xyxy=[round(x0, 1), round(y0, 1), round(x1, 1), round(y1, 1)])) return out def draw_boxes(im, boxes, line=None, font_px=None, fill_alpha=0): """Return a copy of `im` with one rectangle per box and a 'class 0.87' tag in the class colour. Big regions (space_region) are drawn first so smaller elements stay on top.""" im = im.convert('RGB') w, h = im.size line = line or max(2, round(max(w, h) / 900)) font_px = font_px or max(11, round(max(w, h) / 110)) font = _font(font_px) layer = Image.new('RGBA', im.size, (0, 0, 0, 0)) d = ImageDraw.Draw(layer) order = sorted(boxes, key=lambda b: -(b['box_xyxy'][2] - b['box_xyxy'][0]) * (b['box_xyxy'][3] - b['box_xyxy'][1])) for b in order: c = COLORS[b['class_id']] x0, y0, x1, y1 = b['box_xyxy'] if fill_alpha: d.rectangle([x0, y0, x1, y1], fill=(*c, fill_alpha)) d.rectangle([x0, y0, x1, y1], outline=(*c, 255), width=line) for b in order: # labels after all boxes so no rectangle covers a label c = COLORS[b['class_id']] x0, y0, x1, y1 = b['box_xyxy'] text = f"{b['label']} {b['confidence']:.2f}" tx0, ty0, tx1, ty1 = d.textbbox((0, 0), text, font=font) tw, th = tx1 - tx0 + 6, ty1 - ty0 + 4 lx, ly = x0, y0 - th if y0 - th >= 0 else y0 # above the box, or inside it at the top edge lx = min(max(0, lx), w - tw) d.rectangle([lx, ly, lx + tw, ly + th], fill=(*c, 235)) d.text((lx + 3 - tx0, ly + 2 - ty0), text, font=font, fill=(255, 255, 255, 255)) return Image.alpha_composite(im.convert('RGBA'), layer).convert('RGB') def before_after(before, after, gap=None, title_left='Input', title_right='RAQEEB-CCTV detections'): w, h = before.size gap = gap or max(8, w // 100) bar = max(28, round(h / 22)) font = _font(round(bar * 0.6)) canvas = Image.new('RGB', (2 * w + gap, h + bar), 'white') canvas.paste(before.convert('RGB'), (0, bar)) canvas.paste(after, (w + gap, bar)) d = ImageDraw.Draw(canvas) for x, t in ((0, title_left), (w + gap, title_right)): d.text((x + 6, round(bar * 0.15)), t, font=font, fill=(20, 20, 20)) return canvas def legend(classes, width=900): font = _font(20) rows = (len(classes) + 3) // 4 im = Image.new('RGB', (width, 34 * rows + 10), 'white') d = ImageDraw.Draw(im) for k, c in enumerate(classes): x, y = 10 + (k % 4) * (width // 4), 8 + (k // 4) * 34 d.rectangle([x, y + 3, x + 22, y + 25], outline=COLORS[c], width=4) d.text((x + 30, y + 2), NAMES[c], font=font, fill=(20, 20, 20)) return im def main(): ap = argparse.ArgumentParser() ap.add_argument('inputs', nargs='+', type=Path) ap.add_argument('--weights', type=Path, required=True) ap.add_argument('--thresholds', type=Path, required=True) ap.add_argument('--out', type=Path, required=True) ap.add_argument('--device', default='0') ap.add_argument('--imgsz', type=int, default=1024) ap.add_argument('--tile', type=int, default=0, help='0 = whole page (default)') ap.add_argument('--overlap', type=int, default=256) ap.add_argument('--dpi', type=float, default=200) ap.add_argument('--max-side', type=int, default=2400, help='downscale the saved images to this long side') a = ap.parse_args() from ultralytics import YOLO from infer_raqeeb import image_pages, infer_page a.out.mkdir(parents=True, exist_ok=True) config = json.loads(a.thresholds.read_text()) model = YOLO(str(a.weights)) args = SimpleNamespace(tile=a.tile, overlap=a.overlap, imgsz=a.imgsz, device=a.device) for path in a.inputs: for im, meta in image_pages(path, a.dpi, None): dets, _ = infer_page(model, im, config, args) boxes = boxes_from_detections(dets) stem = path.stem + (f"_p{meta['page_number']}" if path.suffix.lower() == '.pdf' else '') s = min(1.0, a.max_side / max(im.size)) small = im.resize((round(im.width * s), round(im.height * s)), Image.LANCZOS) if s < 1 else im sboxes = [dict(b, box_xyxy=[v * s for v in b['box_xyxy']]) for b in boxes] after = draw_boxes(small, sboxes) after.save(a.out / f'{stem}_annotated.jpg', quality=92) before_after(small, after).save(a.out / f'{stem}_before_after.jpg', quality=90) counts = {} for b in boxes: counts[b['label']] = counts.get(b['label'], 0) + 1 (a.out / f'{stem}_boxes.json').write_text(json.dumps(dict( input=str(path), page=meta['page_number'], image_size=list(im.size), counts=counts, boxes=boxes), indent=1)) print(f'{stem}: {len(boxes)} boxes {counts}', flush=True) if __name__ == '__main__': main()