fix: tight polygon 라벨 + eval GT mat 생성 파이프라인 추가

- generate_synthetic.py: plate 전체 box → 글자 실제 좌표 기반 tight polygon으로 변경
  (글자 반복 출력·over-segmentation의 근본 원인 해결)
  gen_type1/2/_gen_two_line 모두 (plate, label_list) 통일 반환
- tools/make_gt_mat.py: test.txt → ICDAR wordBB 포맷 gt_img_N.mat 생성 스크립트 신규
  (E2EMetric seqerr=0.99 고착 문제 해결)
- scripts/run_step1.sh: 데이터 생성 → GT mat 생성 → 학습 3단계로 재구성
  NUM_SAMPLES 환경변수로 데이터 수 제어 가능

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

data_gen/generate_synthetic.py

@@ -89,6 +89,7 @@ class LPGenerator:
         return {k: cv2.resize(v, (w, h)) for k, v in d.items()}
 
     def gen_type1(self):
+        """신형 승용 가로 1줄 (520x110). 글자 실제 위치로 tight polygon 생성."""
         plate = cv2.resize(self.plate_w, (520, 110))
         num = self._resize_dict(self.num_w, 56, 83)
         char = self._resize_dict(self.char_w, 60, 83)
@@ -98,16 +99,20 @@ class LPGenerator:
         e = [random.choice('0123456789') for _ in range(4)]
 
         row, col = 13, 35
+        x0 = col
         for x in d:
             plate[row:row+83, col:col+56] = num[x]; col += 56
         plate[row:row+83, col:col+60] = char[ch]; col += 60 + 36
         for x in e:
             plate[row:row+83, col:col+56] = num[x]; col += 56
+        x1 = col
 
         text = ''.join(d) + HANGUL_CHAR_MAP[ch] + ''.join(e)
-        return plate, text
+        poly = [[x0, row], [x1, row], [x1, row + 83], [x0, row + 83]]
+        return plate, [{"transcription": text, "points": poly}]
 
     def gen_type2(self):
+        """구형 승용 가로 1줄 (355x155). 글자 실제 위치로 tight polygon 생성."""
         plate = cv2.resize(self.plate_w, (355, 155))
         num = self._resize_dict(self.num_w, 45, 83)
         char = self._resize_dict(self.char_w, 49, 70)
@@ -117,17 +122,21 @@ class LPGenerator:
         e = [random.choice('0123456789') for _ in range(4)]
 
         row, col = 46, 10
+        x0 = col
         plate[row:row+83, col:col+45] = num[d[0]]; col += 45
         plate[row:row+83, col:col+45] = num[d[1]]; col += 45
         plate[row+12:row+82, col+2:col+51] = char[ch]; col += 51
         plate[row:row+83, col+2:col+47] = num[e[0]]; col += 47
         for x in e[1:]:
             plate[row:row+83, col:col+45] = num[x]; col += 45
+        x1 = col
 
         text = ''.join(d) + HANGUL_CHAR_MAP[ch] + ''.join(e)
-        return plate, text
+        poly = [[x0, row], [x1, row], [x1, row + 83], [x0, row + 83]]
+        return plate, [{"transcription": text, "points": poly}]
 
     def _gen_two_line(self, plate_bg, num_src, char_src, region_src):
+        """두 줄 LP (336x170). 위·아래 줄 각각 tight polygon 생성."""
         plate = cv2.resize(plate_bg, (336, 170))
         num1 = self._resize_dict(num_src, 44, 60)
         num2 = self._resize_dict(num_src, 64, 90)
@@ -139,48 +148,36 @@ class LPGenerator:
         ch = random.choice(list(HANGUL_CHAR_MAP))
         e = [random.choice('0123456789') for _ in range(4)]
 
+        # 위 줄: region + 숫자2
         row, col = 8, 76
+        tx0 = col
         plate[row:row+60, col:col+88] = region[rkey]; col += 88 + 8
         for x in d:
             plate[row:row+60, col:col+44] = num1[x]; col += 44
+        tx1 = col
+        top_poly = [[tx0, row], [tx1, row], [tx1, row + 60], [tx0, row + 60]]
 
+        # 아래 줄: 한글 + 숫자4
         row, col = 72, 8
+        bx0 = col
         plate[row:row+62, col:col+64] = char[ch]; col += 64
         for x in e:
             plate[row:row+90, col:col+64] = num2[x]; col += 64
+        bx1 = col
+        bot_poly = [[bx0, row], [bx1, row], [bx1, row + 90], [bx0, row + 90]]
 
-        # transcription: 줄 단위로 분리 (PGNet은 두 polygon으로 라벨링하는 게 정석)
         top = REGION_MAP.get(rkey, '?') + ''.join(d)
         bot = HANGUL_CHAR_MAP[ch] + ''.join(e)
-        return plate, top, bot
+        return plate, [
+            {"transcription": top, "points": top_poly},
+            {"transcription": bot, "points": bot_poly},
+        ]
 
     def gen_type3(self):
-        plate, top, bot = self._gen_two_line(self.plate_y, self.num_y, self.char_y, self.region_y_imgs)
-        return plate, top, bot
+        return self._gen_two_line(self.plate_y, self.num_y, self.char_y, self.region_y_imgs)
 
     def gen_type4(self):
-        plate, top, bot = self._gen_two_line(self.plate_g, self.num_g, self.char_g, self.region_g_imgs)
-        return plate, top, bot
-
-
-def make_label_one_line(plate, text):
-    """가로 한 줄 LP — polygon은 plate 전체."""
-    h, w = plate.shape[:2]
-    poly = [[0, 0], [w, 0], [w, h], [0, h]]
-    return [{"transcription": text, "points": poly}]
-
-
-def make_label_two_line(plate, top, bot):
-    """두 줄 LP — 위·아래 두 polygon 으로 분리.
-       위 줄 0~50% (region+num), 아래 줄 50~100% (char+num*4) — 단순 분할."""
-    h, w = plate.shape[:2]
-    mid = h // 2
-    up_poly = [[0, 0], [w, 0], [w, mid], [0, mid]]
-    dn_poly = [[0, mid], [w, mid], [w, h], [0, h]]
-    return [
-        {"transcription": top, "points": up_poly},
-        {"transcription": bot, "points": dn_poly},
-    ]
+        return self._gen_two_line(self.plate_g, self.num_g, self.char_g, self.region_g_imgs)
 
 
 # 한국 도로 LP 분포 추정 (자가용 92% + 영업용 7.5%, 신형 가로 ~98% 등)
@@ -207,10 +204,10 @@ def main():
 
     gen = LPGenerator(Path(args.asset_dir))
     type_funcs = {
-        '1': ('one', gen.gen_type1),
-        '2': ('one', gen.gen_type2),
-        '3': ('two', gen.gen_type3),
-        '4': ('two', gen.gen_type4),
+        '1': gen.gen_type1,
+        '2': gen.gen_type2,
+        '3': gen.gen_type3,
+        '4': gen.gen_type4,
     }
     selected_keys = [t.strip() for t in args.types.split(',') if t.strip() in type_funcs]
     if not selected_keys:
@@ -239,16 +236,11 @@ def main():
         records = []
         for i in range(count):
             idx = random.choices(range(len(chosen)), weights=weights, k=1)[0]
-            kind, fn = chosen[idx]
+            fn = chosen[idx]
             type_count[selected_keys[idx]] += 1
-            if kind == 'one':
-                plate, text = fn()
-                label = make_label_one_line(plate, text)
-                seen_chars.update(text)
-            else:
-                plate, top, bot = fn()
-                label = make_label_two_line(plate, top, bot)
-                seen_chars.update(top); seen_chars.update(bot)
+            plate, label = fn()
+            for entry in label:
+                seen_chars.update(entry["transcription"])
             if not args.no_bright:
                 plate = random_bright(plate)
             fname = f"{i:06d}.jpg"

scripts/run_step1.sh

@@ -14,8 +14,36 @@ set -euo pipefail
 PADDLEOCR_DIR=/workspace/PaddleOCR
 KR_LP_DIR=/workspace/kr_lp_pgnet
 TRAIN_DATA=/workspace/train_data
+SYNTH_DIR="$TRAIN_DATA/kr_lp_synth"
+ASSET_DIR="$KR_LP_DIR/data_gen/Korean-license-plate-Generator"
 LOG="${LOG:-$PADDLEOCR_DIR/output/kr_lp_pgnet/train.log}"
+NUM_SAMPLES="${NUM_SAMPLES:-10000}"
 
+# ── 1. 합성 데이터 생성 ──────────────────────────────────────────────────────
+echo "==========================="
+echo "[1/3] 합성 데이터 생성 (${NUM_SAMPLES}장)"
+echo "  asset: $ASSET_DIR"
+echo "  out:   $SYNTH_DIR"
+echo "==========================="
+
+python3.10 "$KR_LP_DIR/data_gen/generate_synthetic.py" \
+    --asset_dir "$ASSET_DIR" \
+    --out_dir   "$SYNTH_DIR" \
+    --num       "$NUM_SAMPLES" \
+    --dict      "$KR_LP_DIR/dict/kr_lp_dict.txt"
+
+# ── 2. eval GT mat 생성 ─────────────────────────────────────────────────────
+echo "==========================="
+echo "[2/3] eval GT mat 생성"
+echo "  label: $SYNTH_DIR/test/test.txt"
+echo "  out:   $SYNTH_DIR/gt/"
+echo "==========================="
+
+python3.10 "$KR_LP_DIR/tools/make_gt_mat.py" \
+    --label   "$SYNTH_DIR/test/test.txt" \
+    --out_dir "$SYNTH_DIR/gt"
+
+# ── 3. 학습 ─────────────────────────────────────────────────────────────────
 cd "$PADDLEOCR_DIR"
 
 # train_data symlink (config는 ./train_data/kr_lp_synth 사용)
@@ -38,9 +66,9 @@ if [ "${DRY_RUN:-0}" = "1" ]; then
 fi
 
 echo "==========================="
-echo "Step1 학습 시작"
+echo "[3/3] Step1 학습 시작"
 echo "  config:  configs/e2e/kr_lp_pgnet.yml"
-echo "  data:    $TRAIN_DATA/kr_lp_synth/"
+echo "  data:    $SYNTH_DIR/"
 echo "  pretrain: pretrain_models/train_step1/best_accuracy"
 echo "  log:     $LOG"
 echo "  override: ${OVERRIDE[@]}"

tools/make_gt_mat.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""
+PaddleOCR E2EMetric (mode: A)이 요구하는 gt_img_{N}.mat 파일을 생성.
+
+입력: PGDataSet 포맷의 label txt 파일
+  images/000000.jpg\t[{"transcription":"37도1563","points":[[x1,y1],...]}]
+
+출력: gt_mat_dir/gt_img_{1-indexed}.mat (ICDAR2015 wordBB 포맷)
+
+사용법 (서버에서):
+  python3.10 /workspace/kr_lp_pgnet/tools/make_gt_mat.py \
+      --label /workspace/train_data/kr_lp_synth/test/test.txt \
+      --out_dir /workspace/train_data/kr_lp_synth/gt
+"""
+
+import argparse
+import json
+import os
+
+import numpy as np
+import scipy.io as sio
+
+
+def points_to_wordbb(points):
+    """[[x0,y0],[x1,y1],[x2,y2],[x3,y3]] → (2, 4) array (ICDAR wordBB 포맷).
+    axis0: x or y, axis1: corner index (TL, TR, BR, BL)."""
+    pts = np.array(points, dtype=np.float64)  # (4, 2)
+    # (2, 4): row0=x coords, row1=y coords
+    return pts.T  # [[x0..x3], [y0..y3]]
+
+
+def make_txt_cell(text):
+    """ICDAR txt 셀: numpy array of individual chars (MATLAB cell array 호환)."""
+    return np.array([list(text)], dtype=object)
+
+
+def main():
+    p = argparse.ArgumentParser(description=__doc__)
+    p.add_argument("--label", required=True, help="PGDataSet label txt 경로")
+    p.add_argument("--out_dir", required=True, help="gt mat 출력 디렉토리")
+    args = p.parse_args()
+
+    os.makedirs(args.out_dir, exist_ok=True)
+
+    with open(args.label, encoding="utf-8") as f:
+        lines = [ln.rstrip("\n") for ln in f if ln.strip()]
+
+    for img_idx, line in enumerate(lines, start=1):
+        parts = line.split("\t", 1)
+        if len(parts) != 2:
+            print(f"  skip (no tab): {line[:60]}")
+            continue
+
+        try:
+            entries = json.loads(parts[1])
+        except json.JSONDecodeError as ex:
+            print(f"  skip (json err @ {img_idx}): {ex}")
+            continue
+
+        n = len(entries)
+        if n == 0:
+            print(f"  skip (empty label @ {img_idx})")
+            continue
+
+        # wordBB: (2, 4, N) — ICDAR 포맷
+        word_bb = np.zeros((2, 4, n), dtype=np.float64)
+        txt = np.empty((1, n), dtype=object)
+
+        for j, entry in enumerate(entries):
+            word_bb[:, :, j] = points_to_wordbb(entry["points"])
+            txt[0, j] = make_txt_cell(entry["transcription"])
+
+        if n == 1:
+            # scipy.io는 (2,4,1) → (2,4)로 squeeze하는 경우가 있어 명시적 유지
+            word_bb = word_bb.reshape(2, 4, 1)
+
+        mat_path = os.path.join(args.out_dir, f"gt_img_{img_idx}.mat")
+        sio.savemat(mat_path, {"wordBB": word_bb, "txt": txt}, do_compression=True)
+
+    print(f"  gt mat 생성 완료: {len(lines)}개 → {args.out_dir}/gt_img_{{1..{len(lines)}}}.mat")
+
+
+if __name__ == "__main__":
+    main()