feat(detect): 推理侧改为切片，与固定 160 输入一致

训练数据已经是固定边长的片，推理还把整图交给 ultralytics 就会被 letterbox 缩放，
与模型见过的形态不符。三个调用方（可视化、批量推送预测、ML backend）统一改成
切片推理再拼回原图。

- engine/tiling.merge_tiles / tiled_predict：映射回原图、裁掉镜像填充区的框、跨片去重。
  推理框架留在调用方，几何部分可单测。
- engine/detection.nms：标准 NMS，另加按包含比例抑制。缺陷被切割线切开时，相邻片
  只看到一角，给出的小残框与完整框 IoU 很小（面积悬殊），光靠 IoU 压不掉。
- engine/tiled_inference：ultralytics 的接线，内部分批，避免片数多时 OOM。
- 三处调用点加 --tile-size / --tile-overlap（后端为 YOLO_TILE_SIZE / YOLO_TILE_OVERLAP），
  默认 160 / 32；设 0 退回整图推理，仅用于和旧模型对比。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

scripts/push_predictions.py

@@ -40,6 +40,7 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "ml_backend"))
 
 from convert import to_prediction  # noqa: E402
 
+from engine.tiled_inference import load_image, predict_image  # noqa: E402
 from utils import setup_logging  # noqa: E402
 from utils.device import DeviceNotAvailableError, describe_device, resolve_device  # noqa: E402
 
@@ -186,7 +187,12 @@ def main() -> int:
     )
     parser.add_argument("--model-version", default=None, help="缺省按权重文件名与修改时间生成")
     parser.add_argument("--conf", type=float, default=0.25)
-    parser.add_argument("--imgsz", type=int, default=640)
+    parser.add_argument(
+        "--tile-size", type=int, default=160,
+        help="模型输入边长；按它切片推理再拼回，全程不缩放。0=整图推理（会 letterbox）",
+    )
+    parser.add_argument("--tile-overlap", type=int, default=32, help="切片重叠像素")
+    parser.add_argument("--imgsz", type=int, default=640, help="仅 --tile-size 0 时用")
     parser.add_argument("--device", default="auto")
     parser.add_argument("--batch", type=int, default=16, help="每次推理几张")
     parser.add_argument("--chunk", type=int, default=500, help="每次上传几条预测")
@@ -256,12 +262,15 @@ def main() -> int:
         removed = client.delete_predictions(args.project, model_version)
         logger.info("删除同版本（%s）旧预测 %d 条", model_version, removed)
 
-    from PIL import Image
     from ultralytics import YOLO
 
     model = YOLO(str(weights))
     classes = model.names
     device_arg = device.index or 0 if device.type == "cuda" else device.type
+    if args.tile_size > 0:
+        logger.info("切片推理 %dx%d，重叠 %d（不缩放）", args.tile_size, args.tile_size, args.tile_overlap)
+    else:
+        logger.warning("整图推理 imgsz=%d —— ultralytics 会 letterbox 缩放", args.imgsz)
 
     pending: list[dict] = []
     uploaded = with_boxes = total_boxes = 0
@@ -275,35 +284,30 @@ def main() -> int:
         uploaded += len(pending)
         pending = []
 
-    for start in range(0, len(pairs), args.batch):
-        chunk = pairs[start:start + args.batch]
-        results = model.predict(
-            source=[str(path) for _, path in chunk], imgsz=args.imgsz, conf=args.conf,
-            device=device_arg, verbose=False,
+    for processed, (task_id, path) in enumerate(pairs, start=1):
+        image = load_image(path)
+        height, width = image.shape[:2]
+        detections = predict_image(
+            model, image, size=args.tile_size, overlap=args.tile_overlap, conf=args.conf,
+            device=device_arg, batch=args.batch, whole_image_imgsz=args.imgsz,
         )
-        for (task_id, path), result in zip(chunk, results, strict=True):
-            with Image.open(path) as image:
-                width, height = image.size
-            boxes = [
-                (classes[int(class_id)], xyxyn, float(score))
-                for xyxyn, class_id, score in zip(
-                    result.boxes.xyxyn.tolist(), result.boxes.cls.tolist(),
-                    result.boxes.conf.tolist(), strict=True,
-                )
-            ]
-            prediction = to_prediction(boxes, width, height, "label", "image", model_version)
-            prediction["task"] = task_id
-            pending.append(prediction)
-            total_boxes += len(boxes)
-            with_boxes += bool(boxes)
+        boxes = [
+            (classes[box.class_id], (box.x1, box.y1, box.x2, box.y2), box.score)
+            for box in detections
+        ]
+        prediction = to_prediction(boxes, width, height, "label", "image", model_version)
+        prediction["task"] = task_id
+        pending.append(prediction)
+        total_boxes += len(boxes)
+        with_boxes += bool(boxes)
+
         if len(pending) >= args.chunk:
             try:
                 flush()
             except RuntimeError as exc:
                 logger.error("%s", exc)
                 return 1
-        processed = start + len(chunk)
-        if processed % 1000 < args.batch:
+        if processed % 1000 == 0:
             logger.info("处理 [%d/%d]，已上传 %d", processed, len(pairs), uploaded)
     try:
         flush()