File size: 11,308 Bytes
07f5340
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b91b0f5
07f5340
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
"""Build the instance-mask subset: 12 hand-picked frames (polygons below) plus one
mid-clip frame per 10 s clip (polygons in keyframe_polygons.json).

Usage: python 02_gold_subset.py <images_dir> <out_dir> [keyframe_polygons.json]

A frame listed with no objects was checked and contains no instrument or gauze.
The mid-clip frame of clip 66 (source frame 16625) is left out: it shows only
ambiguous fragments at the image edge.

Each object starts as a hand-placed polygon (drawn by Claude on a 60 px
coordinate grid over the original frame). Instrument polygons are then snapped
to image edges with OpenCV GrabCut: the eroded polygon is fixed foreground,
everything outside a dilated polygon is fixed background, the band in between
is decided by colour. The refined mask is kept only if it still agrees with
the hand polygon (IoU >= 0.6); otherwise the polygon itself is used.
Gauze polygons are not refined (woven texture defeats colour models); instrument
masks are subtracted from them so instances never overlap.

Appearance classes (not device identification):
  white_insulated_jawed  jawed instrument with cream/white insulation
  dark_rod               dark cylindrical rod, no visible jaws
  fenestrated_grasper    jawed instrument with an open-window jaw
  gauze                  woven gauze-like material
"""
import json
import sys
from pathlib import Path

import cv2
import numpy as np

img_dir, out_dir = Path(sys.argv[1]), Path(sys.argv[2])
(out_dir / "masks").mkdir(parents=True, exist_ok=True)
(out_dir / "overlays").mkdir(parents=True, exist_ok=True)

W, H, FPS = 720, 576, 25
# frame -> list of (class, confidence, tip or None, polygon)
POLYS = {
    900: [
        ("fenestrated_grasper", "high", (452, 255),
         [(428, 256), (372, 300), (315, 360), (272, 420), (212, 480), (160, 540), (135, 576), (300, 576),
          (292, 540), (322, 480), (358, 420), (398, 360), (455, 300), (480, 265), (462, 250)]),
        ("dark_rod", "medium", (498, 325),
         [(480, 325), (520, 322), (540, 360), (555, 420), (565, 480), (578, 576), (520, 576), (508, 480),
          (495, 420), (480, 360)]),
        ("gauze", "high", None, [(0, 495), (40, 490), (110, 540), (130, 576), (0, 576)]),
    ],
    1800: [
        ("white_insulated_jawed", "high", (405, 205),
         [(0, 435), (120, 330), (240, 245), (310, 205), (375, 168), (402, 165), (415, 195), (412, 240),
          (390, 255), (330, 285), (278, 320), (258, 420), (215, 500), (185, 576), (0, 576)]),
        ("dark_rod", "medium", None,
         [(422, 250), (470, 215), (492, 300), (500, 420), (508, 576), (428, 576), (424, 420)]),
        ("gauze", "medium", None,
         [(150, 440), (280, 370), (340, 400), (360, 500), (390, 576), (185, 576), (215, 500)]),
    ],
    2400: [
        ("white_insulated_jawed", "high", (390, 268),
         [(20, 576), (25, 555), (175, 425), (300, 300), (365, 262), (392, 258), (396, 285), (372, 320),
          (255, 440), (220, 500), (145, 576)]),
        ("dark_rod", "medium", (510, 288),
         [(485, 292), (538, 285), (552, 360), (568, 440), (600, 576), (520, 576), (500, 440), (490, 360)]),
    ],
    3900: [
        ("white_insulated_jawed", "high", (438, 325),
         [(425, 318), (447, 325), (445, 360), (400, 400), (330, 470), (290, 530), (250, 576), (140, 576),
          (230, 490), (300, 420), (360, 370), (400, 335)]),
        ("dark_rod", "low", None,
         [(478, 312), (510, 315), (545, 400), (575, 480), (605, 576), (555, 576), (530, 490), (500, 400)]),
        ("gauze", "high", None,
         [(0, 10), (90, 5), (150, 60), (180, 110), (290, 125), (300, 200), (230, 240), (40, 230), (0, 200)]),
    ],
    4800: [
        ("white_insulated_jawed", "high", (355, 242),
         [(0, 335), (110, 300), (240, 255), (300, 235), (355, 230), (360, 255), (300, 285), (265, 310),
          (150, 370), (40, 430), (0, 445)]),
        ("dark_rod", "low", None, [(290, 425), (335, 420), (345, 480), (340, 535), (300, 535), (285, 480)]),
        ("gauze", "high", None,
         [(0, 20), (60, 10), (120, 60), (230, 90), (270, 140), (240, 200), (180, 250), (60, 310), (0, 330)]),
    ],
    6900: [
        ("white_insulated_jawed", "high", (265, 200),
         [(0, 320), (90, 270), (170, 230), (215, 195), (265, 188), (272, 215), (235, 245), (195, 290),
          (100, 360), (20, 420), (0, 420)]),
        ("dark_rod", "high", (325, 216),
         [(298, 218), (352, 215), (372, 300), (385, 400), (400, 576), (312, 576), (302, 400), (296, 300)]),
        ("gauze", "medium", None, [(425, 440), (480, 415), (520, 470), (540, 540), (470, 560), (430, 520)]),
    ],
    8100: [
        ("white_insulated_jawed", "high", (218, 125),
         [(0, 95), (120, 92), (205, 98), (222, 120), (215, 150), (180, 175), (90, 230), (0, 290)]),
        ("dark_rod", "high", (345, 118),
         [(310, 122), (378, 116), (385, 300), (388, 576), (285, 576), (300, 300)]),
        ("gauze", "medium", None, [(185, 190), (240, 200), (290, 250), (285, 330), (230, 340), (200, 280)]),
    ],
    9300: [
        ("white_insulated_jawed", "high", (372, 335),
         [(345, 335), (380, 328), (385, 360), (340, 430), (300, 490), (320, 576), (180, 576), (215, 480),
          (270, 420), (310, 370)]),
        ("dark_rod", "medium", (415, 347),
         [(392, 350), (440, 345), (455, 400), (470, 480), (480, 576), (405, 576), (395, 480)]),
        ("gauze", "low", None,
         [(70, 320), (120, 315), (140, 380), (170, 440), (190, 520), (120, 560), (90, 470), (70, 390)]),
    ],
    10800: [
        ("dark_rod", "medium", None, [(300, 60), (395, 70), (410, 300), (425, 576), (325, 576), (312, 300)]),
        ("white_insulated_jawed", "low", None,
         [(0, 45), (100, 20), (165, 15), (200, 40), (150, 80), (80, 130), (0, 185)]),
    ],
    12600: [
        ("white_insulated_jawed", "high", (448, 55),
         [(0, 245), (150, 150), (330, 85), (430, 40), (455, 50), (450, 90), (350, 150), (330, 220),
          (250, 330), (120, 430), (0, 470)]),
        ("dark_rod", "low", None, [(592, 395), (615, 392), (650, 520), (620, 530)]),
        ("gauze", "medium", None,
         [(120, 430), (250, 400), (400, 380), (520, 410), (560, 500), (420, 540), (300, 576), (200, 576),
          (130, 480)]),
    ],
    13500: [
        ("white_insulated_jawed", "medium", (135, 165),
         [(0, 165), (60, 150), (110, 140), (140, 150), (138, 200), (130, 255), (60, 335), (0, 375)]),
        ("dark_rod", "medium", (462, 270),
         [(440, 275), (490, 265), (520, 330), (540, 420), (560, 576), (480, 576), (455, 420), (435, 330)]),
        ("gauze", "medium", None,
         [(140, 310), (300, 330), (420, 300), (440, 400), (430, 576), (280, 576), (190, 420), (130, 360)]),
    ],
    15900: [
        ("white_insulated_jawed", "medium", (398, 215),
         [(385, 210), (410, 220), (400, 260), (350, 330), (330, 420), (290, 576), (130, 576), (190, 470),
          (270, 380), (320, 300), (360, 240)]),
        ("dark_rod", "low", None, [(405, 312), (440, 302), (470, 400), (500, 576), (420, 576), (410, 420)]),
    ],
}
if len(sys.argv) > 3:
    for k, objs in json.load(open(sys.argv[3])).items():
        POLYS[int(k)] = [(c, conf, tuple(t) if t else None, [tuple(p) for p in poly]) for c, conf, t, poly in objs]
POLYS = dict(sorted(POLYS.items()))
COLORS = {"white_insulated_jawed": (255, 200, 0), "dark_rod": (0, 165, 255),
          "fenestrated_grasper": (0, 255, 0), "gauze": (255, 0, 255)}
CAT_ID = {"white_insulated_jawed": 1, "dark_rod": 2, "fenestrated_grasper": 3, "gauze": 4}


def poly_mask(poly):
    m = np.zeros((H, W), np.uint8)
    cv2.fillPoly(m, [np.array(poly, np.int32)], 1)
    return m


def refine(img, pm):
    k_in = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15))
    k_out = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (41, 41))
    core, halo = cv2.erode(pm, k_in), cv2.dilate(pm, k_out)
    gc = np.full((H, W), cv2.GC_BGD, np.uint8)
    gc[halo == 1] = cv2.GC_PR_BGD
    gc[pm == 1] = cv2.GC_PR_FGD
    gc[core == 1] = cv2.GC_FGD
    if core.sum() < 50:
        return pm, 1.0, "polygon"
    cv2.grabCut(img, gc, None, np.zeros((1, 65)), np.zeros((1, 65)), 5, cv2.GC_INIT_WITH_MASK)
    out = ((gc == cv2.GC_FGD) | (gc == cv2.GC_PR_FGD)).astype(np.uint8)
    out = cv2.morphologyEx(out, cv2.MORPH_OPEN, cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5)))
    n, lab, stats, _ = cv2.connectedComponentsWithStats(out)
    if n > 1:  # keep the largest component only
        out = (lab == 1 + int(np.argmax(stats[1:, cv2.CC_STAT_AREA]))).astype(np.uint8)
    iou = float((out & pm).sum() / max(1, (out | pm).sum()))
    return (out, iou, "grabcut") if iou >= 0.6 else (pm, iou, "polygon")


coco = {"info": {"description": "instance-mask subset; AI-drafted polygons snapped with GrabCut, reviewed and approved by the contributing physician on 2026-11-17"},
        "images": [], "annotations": [],
        "categories": [{"id": v, "name": k} for k, v in CAT_ID.items()]}
ann_id = 1
for fi, objs in POLYS.items():
    img = cv2.imread(str(img_dir / f"f{fi:05d}.png"))
    assert img is not None and img.shape[:2] == (H, W)
    coco["images"].append({"id": fi, "file_name": f"images/f{fi:05d}.png", "width": W, "height": H,
                           "source_frame_index": fi, "timestamp_seconds": fi / FPS})
    vis = img.copy()
    inst = np.zeros((H, W), np.uint8)
    instruments = np.zeros((H, W), np.uint8)
    for k, (cls, conf, tip, poly) in enumerate(objs, 1):
        pm = poly_mask(poly)
        if cls == "gauze":  # instruments lie in front of gauze: cut them out of the gauze polygon
            mask, iou, method = pm & (1 - instruments), 1.0, "polygon"
        else:
            mask, iou, method = refine(img, pm)
            instruments |= mask
        inst[mask == 1] = k
        if mask.sum() == 0:
            continue
        cnts, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        cnt = max(cnts, key=cv2.contourArea)
        seg = cv2.approxPolyDP(cnt, 1.5, True).reshape(-1).tolist()
        x, y, w, h = cv2.boundingRect(cnt)
        coco["annotations"].append({
            "id": ann_id, "image_id": fi, "category_id": CAT_ID[cls], "iscrowd": 0,
            "bbox": [x, y, w, h], "area": int(mask.sum()), "segmentation": [seg],
            "tip": list(tip) if tip else None, "confidence": conf, "mask_method": method,
            "polygon_vs_refined_iou": round(iou, 3), "instance_index_in_mask_png": k})
        ann_id += 1
        col = COLORS[cls]
        vis[mask == 1] = (0.55 * vis[mask == 1] + 0.45 * np.array(col)).astype(np.uint8)
        cv2.drawContours(vis, [cnt], -1, col, 2)
        cv2.putText(vis, f"{cls} [{conf}/{method}]", (x + 3, max(12, y + 14)), cv2.FONT_HERSHEY_SIMPLEX, 0.4,
                    (255, 255, 255), 1)
        if tip:
            cv2.drawMarker(vis, tip, (0, 0, 255), cv2.MARKER_CROSS, 14, 2)
    cv2.imwrite(str(out_dir / "masks" / f"f{fi:05d}_instances.png"), inst)
    cv2.imwrite(str(out_dir / "overlays" / f"f{fi:05d}_overlay.jpg"), vis, [cv2.IMWRITE_JPEG_QUALITY, 90])

json.dump(coco, open(out_dir / "instances_coco.json", "w"), indent=1)
print("images", len(coco["images"]), "annotations", len(coco["annotations"]))
for a in coco["annotations"]:
    print(a["image_id"], a["category_id"], a["mask_method"], a["polygon_vs_refined_iou"], a["bbox"])