"""字符检测(Detection / End-to-End Spotting)打分。 - Detection:仅看 bbox,IoU > 0.75 视为 TP,包含 [UNK] - Spotting:IoU > 0.75 且字符匹配视为 TP,排除 [UNK] 均输出 per-sample F1。GT bbox 像素单位会被归一化到 0-1000 与模型输出对齐。 """ from __future__ import annotations from ..utils.unk import is_unk_char IOU_THRESH = 0.75 def _iou(a, b) -> float: ax1, ay1, ax2, ay2 = a bx1, by1, bx2, by2 = b if ax2 <= ax1 or ay2 <= ay1 or bx2 <= bx1 or by2 <= by1: return 0.0 ix1, iy1 = max(ax1, bx1), max(ay1, by1) ix2, iy2 = min(ax2, bx2), min(ay2, by2) iw, ih = ix2 - ix1, iy2 - iy1 if iw <= 0 or ih <= 0: return 0.0 inter = iw * ih union = (ax2 - ax1) * (ay2 - ay1) + (bx2 - bx1) * (by2 - by1) - inter return inter / union if union > 0 else 0.0 def _parse_gt(row: dict) -> tuple[list[dict], int, int]: """解析 GT spotting,返回 [{'bbox':[x1,y1,x2,y2],'char':str}, ...] + 图像 W/H。""" sp = row.get("spotting") or [] W = int(row.get("width") or 0) H = int(row.get("height") or 0) items: list[dict] = [] for it in sp: if not isinstance(it, dict): continue ch = it.get("modern_char") if ch is None: ch = it.get("text", "") ch = str(ch or "").strip() bbox = it.get("bbox") if bbox is None: continue x1 = y1 = x2 = y2 = None if isinstance(bbox, dict): try: x1, y1, x2, y2 = (float(bbox[k]) for k in ("x1", "y1", "x2", "y2")) except (KeyError, TypeError, ValueError): continue elif isinstance(bbox, (list, tuple)) and len(bbox) == 4: try: bx = [float(v) for v in bbox] except (TypeError, ValueError): continue if bx[2] < bx[0] or bx[3] < bx[1]: x1, y1, x2, y2 = bx[0], bx[1], bx[0] + bx[2], bx[1] + bx[3] else: if "modern_char" in it: x1, y1, x2, y2 = bx[0], bx[1], bx[0] + bx[2], bx[1] + bx[3] else: x1, y1, x2, y2 = bx else: continue if x2 <= x1 or y2 <= y1: continue items.append({"bbox": [float(x1), float(y1), float(x2), float(y2)], "char": ch}) return items, W, H def _scale_to_1000(items: list[dict], W: int, H: int) -> list[dict]: if W <= 0 or H <= 0: return [] sx, sy = 1000.0 / W, 1000.0 / H return [ {"bbox": [it["bbox"][0] * sx, it["bbox"][1] * sy, it["bbox"][2] * sx, it["bbox"][3] * sy], "char": it["char"]} for it in items ] def _match(preds: list[dict], gts: list[dict], iou_thresh: float, require_char: bool) -> tuple[int, int, int]: if not preds or not gts: return 0, len(preds), len(gts) pairs: list[tuple[float, int, int]] = [] for pi, p in enumerate(preds): for gi, g in enumerate(gts): if require_char and p.get("char", "") != g.get("char", ""): continue iou = _iou(p["bbox"], g["bbox"]) if iou >= iou_thresh: pairs.append((iou, pi, gi)) pairs.sort(key=lambda x: x[0], reverse=True) used_p, used_g, tp = set(), set(), 0 for _, pi, gi in pairs: if pi in used_p or gi in used_g: continue used_p.add(pi) used_g.add(gi) tp += 1 return tp, len(preds) - tp, len(gts) - tp def _f1(tp: int, fp: int, fn: int) -> float: if tp == 0: return 0.0 p = tp / (tp + fp) if (tp + fp) else 0.0 r = tp / (tp + fn) if (tp + fn) else 0.0 return 2 * p * r / (p + r) if (p + r) else 0.0 def judge(extract: dict, row: dict, iou_thresh: float = IOU_THRESH) -> dict: gt_raw, W, H = _parse_gt(row) gts = _scale_to_1000(gt_raw, W, H) preds_src = (extract or {}).get("items") or [] preds: list[dict] = [] for it in preds_src: if not isinstance(it, dict): continue bb = it.get("bbox") if not isinstance(bb, (list, tuple)) or len(bb) != 4: continue try: bb = [float(v) for v in bb] except (TypeError, ValueError): continue if bb[2] <= bb[0] or bb[3] <= bb[1]: continue preds.append({"bbox": bb, "char": str(it.get("char", "") or "").strip()}) det_tp, det_fp, det_fn = _match(preds, gts, iou_thresh, require_char=False) det_f1 = _f1(det_tp, det_fp, det_fn) spot_gts = [it for it in gts if not is_unk_char(it["char"])] spot_preds = [it for it in preds if not is_unk_char(it["char"])] spot_tp, spot_fp, spot_fn = _match(spot_preds, spot_gts, iou_thresh, require_char=True) spot_f1 = _f1(spot_tp, spot_fp, spot_fn) return { "score": spot_f1, "iou_thresh": iou_thresh, "detection_f1": det_f1, "spotting_f1": spot_f1, "detection": {"tp": det_tp, "fp": det_fp, "fn": det_fn}, "spotting": {"tp": spot_tp, "fp": spot_fp, "fn": spot_fn}, "n_pred": len(preds), "n_gt": len(gts), }