added new v11 head

ultralytics/cfg/models/11/yolo11.yaml

@@ -44,4 +44,4 @@ head:
   - [[-1, 10], 1, Concat, [1]] # cat head P5
   - [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)
 
-  - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)
+  - [[16, 19, 22], 1, v11Detect, [nc]] # Detect(P3, P4, P5)

ultralytics/nn/modules/__init__.py

@@ -72,7 +72,7 @@ from .conv import (
     RepConv,
     SpatialAttention,
 )
-from .head import OBB, Classify, Detect, Pose, RTDETRDecoder, Segment, WorldDetect, v10Detect
+from .head import OBB, Classify, Detect, Pose, RTDETRDecoder, Segment, WorldDetect, v10Detect, v11Detect
 from .transformer import (
     AIFI,
     MLP,
@@ -141,6 +141,7 @@ __all__ = (
     "OBB",
     "WorldDetect",
     "v10Detect",
+    "v11Detect",
     "ImagePoolingAttn",
     "ContrastiveHead",
     "BNContrastiveHead",

ultralytics/nn/modules/head.py

@@ -41,14 +41,7 @@ class Detect(nn.Module):
         self.cv2 = nn.ModuleList(
             nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch
         )
-        self.cv3 = nn.ModuleList(
-            nn.Sequential(
-                nn.Sequential(DWConv(x, x, 3), Conv(x, c3, 1)),
-                nn.Sequential(DWConv(c3, c3, 3), Conv(c3, c3, 1)),
-                nn.Conv2d(c3, self.nc, 1),
-            )
-            for x in ch
-        )
+        self.cv3 = nn.ModuleList(nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
         self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()
 
         if self.end2end:
@@ -595,3 +588,35 @@ class v10Detect(Detect):
             for x in ch
         )
         self.one2one_cv3 = copy.deepcopy(self.cv3)
+
+class v11Detect(Detect):
+    """
+    v10 Detection head from https://arxiv.org/pdf/2405.14458.
+
+    Args:
+        nc (int): Number of classes.
+        ch (tuple): Tuple of channel sizes.
+
+    Attributes:
+        max_det (int): Maximum number of detections.
+
+    Methods:
+        __init__(self, nc=80, ch=()): Initializes the v10Detect object.
+        forward(self, x): Performs forward pass of the v10Detect module.
+        bias_init(self): Initializes biases of the Detect module.
+
+    """
+
+    def __init__(self, nc=80, ch=()):
+        """Initializes the v10Detect object with the specified number of classes and input channels."""
+        super().__init__(nc, ch)
+        c3 = max(ch[0], min(self.nc, 100))  # channels
+        # Light cls head
+        self.cv3 = nn.ModuleList(
+                nn.Sequential(
+                    nn.Sequential(DWConv(x, x, 3), Conv(x, c3, 1)),
+                    nn.Sequential(DWConv(c3, c3, 3), Conv(c3, c3, 1)),
+                    nn.Conv2d(c3, self.nc, 1),
+                )
+                for x in ch
+            )

ultralytics/nn/tasks.py

@@ -60,6 +60,7 @@ from ultralytics.nn.modules import (
     Segment,
     WorldDetect,
     v10Detect,
+    v11Detect,
 )
 from ultralytics.utils import DEFAULT_CFG_DICT, DEFAULT_CFG_KEYS, LOGGER, colorstr, emojis, yaml_load
 from ultralytics.utils.checks import check_requirements, check_suffix, check_yaml
@@ -1041,7 +1042,7 @@ def parse_model(d, ch, verbose=True):  # model_dict, input_channels(3)
             args = [ch[f]]
         elif m is Concat:
             c2 = sum(ch[x] for x in f)
-        elif m in {Detect, WorldDetect, Segment, Pose, OBB, ImagePoolingAttn, v10Detect}:
+        elif m in {Detect, WorldDetect, Segment, Pose, OBB, ImagePoolingAttn, v10Detect, v11Detect}:
             args.append([ch[x] for x in f])
             if m is Segment:
                 args[2] = make_divisible(min(args[2], max_channels) * width, 8)