feat(datasets): YOLO 风格标注数据集与几何/像素增强

- yolo: images/ 与同层 labels/ 的路径约定、`class_id cx cy w h` 归一化行解析
  （兼容分割多边形取外接矩形）、data.yaml 读写；缺失或空 txt 即无缺陷样本
- film_defect: 抛出图像、框、图像级标签；label_mode 决定多热向量还是互斥下标
- transforms: 图框同步的 Resize / Letterbox / CenterCrop / RandomResizedCrop /
  翻转 / ColorJitter / GaussianBlur / PerImageStandardize
- target_encoder: 高斯半径与 heatmap / wh / offset / mask 监督信号编码

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

datasets/target_encoder.py

@@ -0,0 +1,111 @@
+"""把框标注编码成 CenterHead 需要的监督信号。"""
+
+from __future__ import annotations
+
+import numpy as np
+import torch
+
+
+def gaussian_radius(height: float, width: float, min_overlap: float = 0.7) -> float:
+    """CenterNet 的半径公式：保证半径内的框与 GT 的 IoU >= min_overlap。"""
+    a1 = 1.0
+    b1 = height + width
+    c1 = width * height * (1 - min_overlap) / (1 + min_overlap)
+    r1 = (b1 - np.sqrt(max(b1**2 - 4 * a1 * c1, 0.0))) / (2 * a1)
+
+    a2 = 4.0
+    b2 = 2 * (height + width)
+    c2 = (1 - min_overlap) * width * height
+    r2 = (b2 - np.sqrt(max(b2**2 - 4 * a2 * c2, 0.0))) / (2 * a2)
+
+    a3 = 4 * min_overlap
+    b3 = -2 * min_overlap * (height + width)
+    c3 = (min_overlap - 1) * width * height
+    r3 = (-b3 + np.sqrt(max(b3**2 - 4 * a3 * c3, 0.0))) / (2 * a3)
+
+    return max(0.0, min(r1, r2, r3))
+
+
+def draw_gaussian(heatmap: np.ndarray, center: tuple[int, int], radius: int) -> None:
+    """在 heatmap 上叠加二维高斯（取 max，避免相邻目标互相削弱）。"""
+    diameter = 2 * radius + 1
+    sigma = diameter / 6.0
+    grid = np.arange(-radius, radius + 1, dtype=np.float32)
+    gaussian = np.exp(-(grid[None, :] ** 2 + grid[:, None] ** 2) / (2 * sigma**2))
+
+    cx, cy = center
+    height, width = heatmap.shape
+    left, right = min(cx, radius), min(width - cx, radius + 1)
+    top, bottom = min(cy, radius), min(height - cy, radius + 1)
+    if right <= -left or bottom <= -top:
+        return
+
+    masked_heatmap = heatmap[cy - top : cy + bottom, cx - left : cx + right]
+    masked_gaussian = gaussian[radius - top : radius + bottom, radius - left : radius + right]
+    np.maximum(masked_heatmap, masked_gaussian, out=masked_heatmap)
+
+
+class CenterTargetEncoder:
+    """生成 heatmap / wh / offset / mask 四个监督张量。
+
+    Args:
+        num_classes: 类别数。
+        stride: 模型输出相对输入的下采样倍数。
+        max_objects: 单图最多编码的目标数。
+        min_overlap: 高斯半径的 IoU 阈值。
+    """
+
+    def __init__(
+        self,
+        num_classes: int,
+        stride: int,
+        max_objects: int = 64,
+        min_overlap: float = 0.7,
+    ) -> None:
+        self.num_classes = num_classes
+        self.stride = stride
+        self.max_objects = max_objects
+        self.min_overlap = min_overlap
+
+    def __call__(
+        self,
+        boxes: np.ndarray,
+        labels: np.ndarray,
+        image_size: tuple[int, int],
+    ) -> dict[str, torch.Tensor]:
+        img_h, img_w = image_size
+        out_h, out_w = img_h // self.stride, img_w // self.stride
+
+        heatmap = np.zeros((self.num_classes, out_h, out_w), dtype=np.float32)
+        wh = np.zeros((2, out_h, out_w), dtype=np.float32)
+        offset = np.zeros((2, out_h, out_w), dtype=np.float32)
+        mask = np.zeros((out_h, out_w), dtype=np.float32)
+
+        count = 0
+        for box, label in zip(np.asarray(boxes, dtype=np.float32), np.asarray(labels), strict=True):
+            if count >= self.max_objects:
+                break
+            x1, y1, x2, y2 = box / self.stride
+            box_w, box_h = x2 - x1, y2 - y1
+            if box_w <= 0 or box_h <= 0:
+                continue
+
+            cx, cy = (x1 + x2) / 2, (y1 + y2) / 2
+            cx_int, cy_int = int(cx), int(cy)
+            if not (0 <= cx_int < out_w and 0 <= cy_int < out_h):
+                continue
+
+            radius = max(1, int(gaussian_radius(box_h, box_w, self.min_overlap)))
+            draw_gaussian(heatmap[int(label)], (cx_int, cy_int), radius)
+
+            wh[:, cy_int, cx_int] = (box_w, box_h)
+            offset[:, cy_int, cx_int] = (cx - cx_int, cy - cy_int)
+            mask[cy_int, cx_int] = 1.0
+            count += 1
+
+        return {
+            "heatmap": torch.from_numpy(heatmap),
+            "wh": torch.from_numpy(wh),
+            "offset": torch.from_numpy(offset),
+            "mask": torch.from_numpy(mask),
+        }