File size: 6,968 Bytes
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
"""Re-segment the hand-prompted frames with SAM 3.1 (runs on a CUDA machine).

Usage: python 06_sam31_refine_frames.py <images_dir> <instances_coco.json> <checkpoint.pt> <out_dir> [limit]

Every instance in instances_coco.json (AI-drafted polygon, GrabCut-snapped) is
turned into three positive click prompts placed deep inside its mask, one per
third of the object's long axis. SAM 3.1 (official facebookresearch/sam3 code,
checkpoint sam3.1_multiplex.pt) returns one mask per instance. Frames are fed
one at a time, each as its own one-image session; nothing is propagated between
frames here.

Outputs: sam31_masks/<frame>_instances.png, sam31_instances.json (per-instance
IoU against the draft mask) and overlays.
"""
import json
import shutil
import sys
from pathlib import Path

import cv2
import numpy as np
import torch

img_dir, coco_path, ckpt, out_dir = Path(sys.argv[1]), sys.argv[2], sys.argv[3], Path(sys.argv[4])
limit = int(sys.argv[5]) if len(sys.argv) > 5 else None
coco = json.load(open(coco_path))
cat = {c["id"]: c["name"] for c in coco["categories"]}
images = sorted(coco["images"], key=lambda i: i["id"])[:limit]
H, W = 576, 720

jpg_dir = out_dir / "frames_jpg"
if jpg_dir.exists():
    shutil.rmtree(jpg_dir)
jpg_dir.mkdir(parents=True)
(out_dir / "sam31_masks").mkdir(exist_ok=True)
(out_dir / "sam31_overlays").mkdir(exist_ok=True)
for k, im in enumerate(images):
    (jpg_dir / f"{k:05d}").mkdir()
    cv2.imwrite(str(jpg_dir / f"{k:05d}" / "00000.jpg"), cv2.imread(str(img_dir / Path(im["file_name"]).name)),
                [cv2.IMWRITE_JPEG_QUALITY, 98])


def draft_mask(a):
    m = np.zeros((H, W), np.uint8)
    for seg in a["segmentation"]:
        cv2.fillPoly(m, [np.array(seg, np.int32).reshape(-1, 2)], 1)
    return m


def click_points(m, n=3):
    """n clicks: deepest interior point in each third of the object's long axis."""
    dt = cv2.distanceTransform(m, cv2.DIST_L2, 5)
    ys, xs = np.nonzero(m)
    pts = np.stack([xs, ys], 1).astype(np.float32)
    axis = cv2.PCACompute(pts, mean=None)[1][0]
    proj = pts @ axis
    edges = np.quantile(proj, np.linspace(0, 1, n + 1))
    out = []
    for i in range(n):
        sel = (proj >= edges[i]) & (proj <= edges[i + 1])
        if sel.any():
            j = np.argmax(dt[ys[sel], xs[sel]])
            out.append([int(xs[sel][j]), int(ys[sel][j])])
    return out


from sam3.model_builder import build_sam3_multiplex_video_predictor

predictor = build_sam3_multiplex_video_predictor(checkpoint_path=ckpt, use_fa3=False, async_loading_frames=False)
# Upstream start_session passes offload_state_to_cpu, which the SAM 3.1 multiplex
# init_state does not accept (TypeError at repo HEAD of 2026-11-15). Drop that one
# keyword; no model code is changed.
_init_state = predictor.model.init_state
predictor.model.init_state = lambda **kw: _init_state(**{k: v for k, v in kw.items() if k != "offload_state_to_cpu"})
THRESH = 0.0  # keep every returned mask; the score is recorded per instance instead

COLORS = {"white_insulated_jawed": (255, 200, 0), "dark_rod": (0, 165, 255), "fenestrated_grasper": (0, 255, 0),
          "gauze": (255, 0, 255)}
results = []
for k, im in enumerate(images):
    anns = [a for a in coco["annotations"] if a["image_id"] == im["id"]]
    frame = cv2.imread(str(img_dir / Path(im["file_name"]).name))
    vis, inst = frame.copy(), np.zeros((H, W), np.uint8)
    sid = predictor.handle_request(dict(type="start_session", resource_path=str(jpg_dir / f"{k:05d}")))["session_id"]
    clicks, got = {}, {}

    def collect(out):
        ids = [int(i) for i in np.asarray(out["out_obj_ids"]).reshape(-1)]
        masks, probs = np.asarray(out["out_binary_masks"]), np.asarray(out["out_probs"]).reshape(-1)
        for j, oid in enumerate(ids):
            got[oid] = (masks[j].astype(np.uint8).reshape(masks.shape[-2], masks.shape[-1]), float(probs[j]))

    for n, a in enumerate(anns):
        dm = draft_mask(a)
        oid = a["instance_index_in_mask_png"]
        clicks[oid] = click_points(dm)
        # The reply to the very first add_prompt of a session comes back empty (observed
        # with repo HEAD of 2026-11-15); sending the same clicks again returns the mask.
        for _ in range(2 if n == 0 else 1):
            resp = predictor.handle_request(dict(
                type="add_prompt", session_id=sid, frame_index=0, obj_id=oid, output_prob_thresh=THRESH,
                points=[[x / W, y / H] for x, y in clicks[oid]], point_labels=[1] * len(clicks[oid])))
            collect(resp["outputs"])
    predictor.handle_request(dict(type="close_session", session_id=sid))

    # paint gauze first so instruments, which lie in front of it, win overlapping pixels
    for a in sorted(anns, key=lambda a: cat[a["category_id"]] != "gauze"):
        dm, oid = draft_mask(a), a["instance_index_in_mask_png"]
        pts = clicks[oid]
        base = {"image_id": im["id"], "annotation_id": a["id"], "category": cat[a["category_id"]],
                "draft_confidence": a["confidence"], "clicks": pts, "instance_index_in_mask_png": oid}
        if oid not in got or got[oid][0].sum() == 0:
            results.append({**base, "status": "no_mask_returned"})
            continue
        m, prob = got[oid]
        if m.shape != (H, W):
            m = cv2.resize(m, (W, H), interpolation=cv2.INTER_NEAREST)
        iou = float((m & dm).sum() / max(1, (m | dm).sum()))
        inst[m == 1] = oid
        ys, xs = np.nonzero(m)
        results.append({**base, "status": "ok", "sam31_score": round(prob, 3), "iou_vs_draft": round(iou, 3),
                        "area": int(m.sum()), "draft_area": int(dm.sum()),
                        "bbox": [int(xs.min()), int(ys.min()), int(xs.max() - xs.min() + 1), int(ys.max() - ys.min() + 1)]})
        col = COLORS[cat[a["category_id"]]]
        vis[m == 1] = (0.55 * vis[m == 1] + 0.45 * np.array(col)).astype(np.uint8)
        cnts, _ = cv2.findContours(m, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        cv2.drawContours(vis, cnts, -1, col, 2)
        dc, _ = cv2.findContours(dm, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        cv2.drawContours(vis, dc, -1, (255, 255, 255), 1)          # draft outline in white
        for x, y in pts:
            cv2.circle(vis, (x, y), 4, (0, 0, 255), -1)
    cv2.imwrite(str(out_dir / "sam31_masks" / f"f{im['id']:05d}_instances.png"), inst)
    cv2.imwrite(str(out_dir / "sam31_overlays" / f"f{im['id']:05d}_sam31.jpg"), vis, [cv2.IMWRITE_JPEG_QUALITY, 90])

json.dump({"model": "SAM 3.1 (sam3.1_multiplex.pt)", "prompt": "3 positive clicks per instance from the draft mask",
           "instances": results}, open(out_dir / "sam31_instances.json", "w"), indent=1)
ok = [r for r in results if r["status"] == "ok"]
ious = np.array([r["iou_vs_draft"] for r in ok])
print("instances", len(results), "ok", len(ok), "IoU vs draft: median", round(float(np.median(ious)), 3),
      "mean", round(float(ious.mean()), 3), "<0.5:", int((ious < 0.5).sum()))