File size: 8,306 Bytes
8021aca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
"""
tiled_yolo_softnms.py
Tiled inference + class-wise Soft-NMS for YOLO (Ultralytics).
- Runs YOLO on overlapping tiles to boost recall on small symbols.
- Maps all tile detections back to full-image coords.
- Fuses duplicates with Soft-NMS per class.

Usage
-----
from ultralytics import YOLO
import cv2

model = YOLO("best.pt")  # your YOLO v12/v11/v8 checkpoint
img = cv2.imread("example.jpg")[:, :, ::-1]  # BGR->RGB (optional; YOLO accepts BGR too)

out = detect_tiled_softnms(
    model, img,
    tile_size=1024, overlap=0.25,
    per_tile_conf=0.2, per_tile_iou=0.7,
    softnms_iou=0.55, softnms_method="linear", softnms_sigma=0.5,
    final_conf=0.25, device=None, imgsz=None
)

# Access results
xyxy = out["xyxy"]
conf = out["conf"]
cls  = out["cls"]
annot = draw_detections(img.copy(), xyxy, conf, cls, out["names"])
cv2.imwrite("annotated.jpg", annot[:, :, ::-1])  # RGB->BGR for writing
"""

from typing import List, Tuple, Dict, Optional
import numpy as np
import cv2

# ---------------------------
# Utilities
# ---------------------------

def make_overlapping_tiles(H: int, W: int, tile: int, overlap: float) -> List[Tuple[int, int, int, int]]:
    """Return list of (x0, y0, x1, y1) tile boxes covering the image with given overlap."""
    assert 0.0 <= overlap < 1.0
    stride = max(1, int(tile * (1.0 - overlap)))
    xs = list(range(0, max(W - tile, 0) + 1, stride))
    ys = list(range(0, max(H - tile, 0) + 1, stride))
    if xs[-1] + tile < W:
        xs.append(W - tile)
    if ys[-1] + tile < H:
        ys.append(H - tile)
    tiles = []
    for y in ys:
        for x in xs:
            x0, y0 = max(0, x), max(0, y)
            x1, y1 = min(W, x0 + tile), min(H, y0 + tile)
            tiles.append((x0, y0, x1, y1))
    return tiles

def iou_xyxy(a: np.ndarray, b: np.ndarray) -> np.ndarray:
    """IoU between one box a (4,x) and many boxes b (N,4)."""
    xx1 = np.maximum(a[0], b[:, 0])
    yy1 = np.maximum(a[1], b[:, 1])
    xx2 = np.minimum(a[2], b[:, 2])
    yy2 = np.minimum(a[3], b[:, 3])
    inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
    area_a = (a[2]-a[0]) * (a[3]-a[1])
    area_b = (b[:, 2]-b[:, 0]) * (b[:, 3]-b[:, 1])
    union = np.maximum(1e-9, area_a + area_b - inter)
    return inter / union

def soft_nms_classwise(
    boxes: np.ndarray, scores: np.ndarray, classes: np.ndarray,
    iou_thr: float = 0.55, method: str = "linear", sigma: float = 0.5,
    score_thresh: float = 1e-3, max_det: Optional[int] = None
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
    """
    Soft-NMS per class.
    boxes: (N,4), scores:(N, x), classes:(N, x)
    Returns filtered (boxes, scores, classes).
    """
    keep_boxes, keep_scores, keep_classes = [], [], []
    for c in np.unique(classes):
        m = classes == c
        b = boxes[m].astype(np.float32).copy()
        s = scores[m].astype(np.float32).copy()
        idxs = np.arange(b.shape[0])

        kept = []
        while len(idxs):
            i = idxs[np.argmax(s[idxs])]
            M = b[i].copy()
            Ms = s[i].copy()
            kept.append(i)

            idxs = idxs[idxs != i]
            if len(idxs) == 0:
                break
            ious = iou_xyxy(M, b[idxs])
            if method == "linear":
                decay = np.where(ious > iou_thr, 1.0 - ious, 1.0)
                s[idxs] *= decay
            elif method == "gaussian":
                s[idxs] *= np.exp(-(ious ** 2) / sigma)
            elif method == "hard":
                # standard NMS behaviour
                idxs = idxs[ious <= iou_thr]
            else:
                raise ValueError("method must be 'linear', 'gaussian', or 'hard'")

            # prune very low scores
            idxs = idxs[s[idxs] >= score_thresh]

        if kept:
            kb, ks = b[kept], s[kept]
            order = np.argsort(-ks)
            kb, ks = kb[order], ks[order]
            kc = np.full(len(ks), c, dtype=classes.dtype)
            keep_boxes.append(kb)
            keep_scores.append(ks)
            keep_classes.append(kc)

    if not keep_boxes:
        return (np.zeros((0, 4), dtype=np.float32),
                np.zeros((0,), dtype=np.float32),
                np.zeros((0,), dtype=classes.dtype))

    B = np.concatenate(keep_boxes, axis=0)
    S = np.concatenate(keep_scores, axis=0)
    C = np.concatenate(keep_classes, axis=0)

    order = np.argsort(-S)
    if max_det is not None:
        order = order[:max_det]
    return B[order], S[order], C[order]

def draw_detections(img: np.ndarray, boxes: np.ndarray, scores: np.ndarray, classes: np.ndarray, names: Dict[int, str]) -> np.ndarray:
    """Simple visualizer (RGB in, RGB out)."""
    for (x1, y1, x2, y2), sc, cl in zip(boxes.astype(int), scores, classes.astype(int)):
        label = f"{names.get(cl, str(cl))} {sc:.2f}"
        cv2.rectangle(img, (x1, y1), (x2, y2), (0, 180, 255), 2)
        (tw, th), bl = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
        cv2.rectangle(img, (x1, y1 - th - 6), (x1 + tw + 4, y1), (0, 180, 255), -1)
        cv2.putText(img, label, (x1 + 2, y1 - 4), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 0), 2, cv2.LINE_AA)
    return img

# ---------------------------
# Main tiled inference
# ---------------------------

def detect_tiled_softnms(
    model, image: np.ndarray,
    tile_size: int = 1024, overlap: float = 0.25,
    per_tile_conf: float = 0.25, per_tile_iou: float = 0.7,
    softnms_iou: float = 0.55, softnms_method: str = "linear", softnms_sigma: float = 0.5,
    final_conf: float = 0.25, max_det: int = 3000,
    device: Optional[str] = None, imgsz: Optional[int] = None,
    class_agnostic_nms: bool = False
) -> Dict[str, np.ndarray]:
    """
    Run YOLO on overlapping tiles, then fuse globally with class-wise Soft-NMS.
    Returns dict: {"xyxy","conf","cls","names"}.
    """
    assert image.ndim == 3, "image must be HxWx3"
    H, W = image.shape[:2]
    names = getattr(model, "names", {i: str(i) for i in range(1000)})

    tiles = make_overlapping_tiles(H, W, tile=tile_size, overlap=overlap)

    all_boxes, all_scores, all_classes = [], [], []

    for (x0, y0, x1, y1) in tiles:
        tile = image[y0:y1, x0:x1]
        # Ultralytics returns boxes in original tile coords (pre-letterbox)
        results = model.predict(
            source=tile,
            conf=per_tile_conf,
            iou=per_tile_iou,
            imgsz=imgsz,        # None -> model default
            device=device,
            verbose=False
        )

        if not results:
            continue

        r = results[0]
        if r.boxes is None or r.boxes.shape[0] == 0:
            continue

        b = r.boxes.xyxy.cpu().numpy()
        s = r.boxes.conf.cpu().numpy()
        c = r.boxes.cls.cpu().numpy().astype(int)

        # Map to full-image coordinates
        b[:, [0, 2]] += x0
        b[:, [1, 3]] += y0

        # Clip
        b[:, 0] = np.clip(b[:, 0], 0, W - 1)
        b[:, 1] = np.clip(b[:, 1], 0, H - 1)
        b[:, 2] = np.clip(b[:, 2], 0, W - 1)
        b[:, 3] = np.clip(b[:, 3], 0, H - 1)

        # Filter degenerate boxes
        valid = (b[:, 2] > b[:, 0]) & (b[:, 3] > b[:, 1])
        if not np.any(valid):
            continue
        all_boxes.append(b[valid])
        all_scores.append(s[valid])
        all_classes.append(c[valid])

    if not all_boxes:
        return {"xyxy": np.zeros((0, 4), dtype=np.float32),
                "conf": np.zeros((0,), dtype=np.float32),
                "cls": np.zeros((0,), dtype=np.int32),
                "names": names}

    boxes = np.concatenate(all_boxes, axis=0).astype(np.float32)
    scores = np.concatenate(all_scores, axis=0).astype(np.float32)
    classes = np.concatenate(all_classes, axis=0).astype(np.int32)

    # Global fusion: class-wise Soft-NMS or class-agnostic if chosen
    if class_agnostic_nms:
        classes = np.zeros_like(classes)

    boxes, scores, classes = soft_nms_classwise(
        boxes, scores, classes,
        iou_thr=softnms_iou,
        method=softnms_method,
        sigma=softnms_sigma,
        score_thresh=1e-3,
        max_det=max_det
    )

    # Final confidence gate
    keep = scores >= final_conf
    boxes, scores, classes = boxes[keep], scores[keep], classes[keep]

    return {"xyxy": boxes, "conf": scores, "cls": classes, "names": names}