1
diff --git a/ultralytics/data/base.py b/ultralytics/data/base.py
2
--- a/ultralytics/data/base.py
3
+++ b/ultralytics/data/base.py
4
@@ -85,6 +85,7 @@ class BaseDataset(Dataset):
5
         classes: list[int] | None = None,
6
         fraction: float = 1.0,
7
         channels: int = 3,
8
+        ir_dir: str = "images_ir",
9
     ):
10
         """Initialize BaseDataset with given configuration and options."""
11
         super().__init__()
12
@@ -115,6 +117,14 @@ class BaseDataset(Dataset):
13
         self.channels = channels
14
         self.cv2_flag = cv2.IMREAD_GRAYSCALE if channels == 1 else cv2.IMREAD_COLOR
15
         self.im_files = self.get_img_files(self.img_path)
16
+        # Dual-stream: auto-infer IR image paths when channels > 3 (e.g. 6 for RGB+IR)
17
+        self.ir_files = []
18
+        self.ir_dir = ir_dir
19
+        if self.channels > 3:
20
+            self.ir_files = [
21
+                f.replace(os.sep + "images" + os.sep, os.sep + self.ir_dir + os.sep) for f in self.im_files
22
+            ]
23
+            self._verify_multimodal_files()
24
         self.labels = self.get_labels()
25
         self.update_labels(include_class=classes)
26
         self.ni = len(self.labels)

1
diff --git a/ultralytics/data/utils.py b/ultralytics/data/utils.py
2
--- a/ultralytics/data/utils.py
3
+++ b/ultralytics/data/utils.py
4
@@ -60,7 +60,15 @@ def img2label_paths(img_paths):
5
     """Convert image paths to label paths."""
6
     sa, sb = f"{os.sep}images{os.sep}", f"{os.sep}labels{os.sep}"
7
-    return [sb.join(x.rsplit(sa, 1)).rsplit(".", 1)[0] + ".txt" for x in img_paths]
8
+    res = []
9
+    for x in img_paths:
10
+        if sa in x:
11
+            res.append(sb.join(x.rsplit(sa, 1)).rsplit(".", 1)[0] + ".txt")
12
+        elif f"{os.sep}images_ir{os.sep}" in x:
13
+            res.append(sb.join(x.rsplit(f"{os.sep}images_ir{os.sep}", 1)).rsplit(".", 1)[0] + ".txt")
14
+        else:
15
+            res.append(x.rsplit(".", 1)[0] + ".txt")
16
+    return res

1
diff --git a/ultralytics/data/loaders.py b/ultralytics/data/loaders.py
2
--- a/ultralytics/data/loaders.py
3
+++ b/ultralytics/data/loaders.py
4
@@ -385,6 +385,7 @@ class LoadImagesAndVideos:
5
         self.vid_stride = vid_stride
6
         self.bs = batch
7
         self.cv2_flag = cv2.IMREAD_GRAYSCALE if channels == 1 else cv2.IMREAD_COLOR
8
+        self.channels = channels  # Dual-stream: save for IR merge check
9
@@ -445,6 +446,14 @@ class LoadImagesAndVideos:
10
                 self.mode = "image"
11
                 im0 = imread(path, flags=self.cv2_flag)  # BGR
12
+                # Dual-stream: merge IR image when channels > 3
13
+                if getattr(self, 'channels', 3) > 3 and im0 is not None:
14
+                    ir_path = path.replace(os.sep + "images" + os.sep, os.sep + "images_ir" + os.sep)
15
+                    if not os.path.exists(ir_path):
16
+                        ir_path = os.path.splitext(path)[0] + '_ir' + os.path.splitext(path)[1]
17
+                    im_ir = imread(ir_path, flags=self.cv2_flag)
18
+                    if im_ir is not None:
19
+                        im0 = cv2.merge((*cv2.split(im0), *cv2.split(im_ir)))  # RGB(3ch) + IR(3ch) = 6ch

1
diff --git a/ultralytics/nn/modules/block.py b/ultralytics/nn/modules/block.py
2
--- a/ultralytics/nn/modules/block.py
3
+++ b/ultralytics/nn/modules/block.py
4
@@ -52,6 +52,8 @@ __all__ = (
5
     "ResNetLayer",
6
     "SCDown",
7
     "TorchVision",
8
+    "IN",
9
+    "Multiin",
10
 )
11

12
+class IN(nn.Module):
13
+    """Identity node for indexing in YAML topology.
14
+
15
+    This module acts as a pass-through placeholder, useful for creating explicit
16
+    index reference points in the YAML model definition when building dual-stream
17
+    or multi-branch architectures.
18
+    """
19
+
20
+    def __init__(self):
21
+        super().__init__()
22
+
23
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
24
+        return x
25
+
26
+
27
+class Multiin(nn.Module):
28
+    """Split multi-channel input into sub-streams along the channel dimension.
29
+
30
+    Used in multi-stream architectures to separate a composite input tensor into
31
+    independent tensors for parallel processing.
32
+
33
+    Attributes:
34
+        out (int): Which sub-stream to output (1-indexed).
35
+        streams (int): Total number of sub-streams the tensor is composed of.
36
+    """
37
+
38
+    def __init__(self, out: int = 1, streams: int = 2):
39
+        super().__init__()
40
+        self.out = out
41
+        self.streams = streams
42
+
43
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
44
+        c = x.shape[1] // self.streams
45
+        return x[:, (self.out - 1) * c : self.out * c]

1
from ultralytics.nn.modules import (
2
    ...
3
    IN,
4
    Index,
5
    LRPCHead,
6
    Multiin,
7
    Pose,
8
    ...
9
    Add,
10
    CAFF,
11
    CDC,
12
    CSSA,
13
    ChannelSwitching,
14
    CombinedFusionBlocks,
15
    ConcatFusion,
16
    ECABlock,
17
    ICFusion,
18
    ITFuse,
19
    MaxFusion,
20
    SFEG,
21
    SpatialAttention,
22
    TFusion,
23
    TransformerFusionBlock,
24
)

1
elif m in fusion_modules:
2
    if m is TransformerFusionBlock:
3
        c2 = ch[f[0]]  # 输出通道=输入通道
4
        args = [c2, *args[1:]]
5
    elif m in [CDC]:
6
        c2 = ch[f[0]]
7
        args = [c2, args[1], args[2]]
8
    elif m in {SFEG, ConcatFusion, CAFF, ICFusion, TFusion}:
9
        c2 = ch[f[0]]  # output channels = input channels
10
        args = [c2, *args]  # prepend in_channels
11
    else:
12
        # No-arg fusion modules: Add, MaxFusion, CSSA, etc.
13
        for x in f:
14
            if x >= -1:
15
                c2 = ch[x]
16
                continue

1
M3FD/
2
├── images/
3
│   ├── train/
4
│   │   ├── 00001.png  (RGB)
5
│   │   └── ...
6
│   └── val/
7
├── images_ir/
8
│   ├── train/
9
│   │   ├── 00001.png  (IR)
10
│   │   └── ...
11
│   └── val/
12
└── labels/
13
    ├── train/
14
    │   ├── 00001.txt
15
    │   └── ...
16
    └── val/

1
nc: 80
2
end2end: True
3
reg_max: 1
4
scales:
5
  n: [0.50, 0.25, 1024]
6
  s: [0.50, 0.50, 1024]
7
  m: [0.50, 1.00, 512]
8
  l: [1.00, 1.00, 512]
9
  x: [1.00, 1.50, 512]
10

11
backbone:
12
  - [-1, 1, Conv, [64, 3, 2]]          # 0-P1/2
13
  - [-1, 1, Conv, [128, 3, 2]]         # 1-P2/4
14
  - [-1, 2, C3k2, [256, False, 0.25]]  # 2
15
  - [-1, 1, Conv, [256, 3, 2]]         # 3-P3/8
16
  - [-1, 2, C3k2, [512, False, 0.25]]  # 4
17
  - [-1, 1, Conv, [512, 3, 2]]         # 5-P4/16
18
  - [-1, 2, C3k2, [512, True]]         # 6
19
  - [-1, 1, Conv, [1024, 3, 2]]        # 7-P5/32
20
  - [-1, 2, C3k2, [1024, True]]        # 8
21
  - [-1, 1, SPPF, [1024, 5, 3, True]]  # 9
22
  - [-1, 2, C2PSA, [1024]]             # 10
23

24
head:
25
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
26
  - [[-1, 6], 1, Concat, [1]]
27
  - [-1, 2, C3k2, [512, True]]         # 13
28
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
29
  - [[-1, 4], 1, Concat, [1]]
30
  - [-1, 2, C3k2, [256, True]]         # 16 (P3/8)
31
  - [-1, 1, Conv, [256, 3, 2]]
32
  - [[-1, 13], 1, Concat, [1]]
33
  - [-1, 2, C3k2, [512, True]]         # 19 (P4/16)
34
  - [-1, 1, Conv, [512, 3, 2]]
35
  - [[-1, 10], 1, Concat, [1]]
36
  - [-1, 1, C3k2, [1024, True, 0.5, True]]  # 22 (P5/32)
37
  - [[16, 19, 22], 1, Detect, [nc]]

1
# yolo26-dual-late.yaml（结构概要，完整版约 99 行）
2
ch: 6
3
nc: 80
4

5
backbone:
6
  # --- Input split ---
7
  - [-1, 1, IN, []]            # 0
8
  - [-1, 1, Multiin, [1]]      # 1  RGB
9
  - [-2, 1, Multiin, [2]]      # 2  IR
10

11
  # --- 完整的并行双骨干 ---
12
  # Stream A (RGB): P1→P2→P3→P4→P5 (nodes 3-20)
13
  # Stream B (IR):  P1→P2→P3→P4→P5 (nodes 3-20, 但独立权重)
14

15
head:
16
  # --- 在每个尺度级别融合两条流 ---
17
  # P5: Upsample(A) + Upsample(B) → Concat(A,P4) + Concat(B,P4) → ...
18
  # 每一层都有 A/B 两个分支
19
  # 最终从 6 个特征图检测：
20
  - [[31, 32, 37, 38, 43, 44], 1, Detect, [nc]]

1
diff --git a/ultralytics/models/yolo/detect/val.py b/ultralytics/models/yolo/detect/val.py
2
# 在 plot_val_samples 中添加 IR 标签绘制
3
        if batch["img"].shape[1] > 3:
4
            ir_batch = {**batch, "img": batch["img"][:, 3:]}
5
            plot_images(
6
                labels=ir_batch,
7
                paths=batch["im_file"],
8
                fname=self.save_dir / f"val_batch_ir{ni}_labels.jpg",
9
                names=self.names,
10
                on_plot=self.on_plot,
11
            )
12

13
# 在 plot_predictions 中添加 IR 预测绘制
14
        if batch["img"].shape[1] > 3:
15
            plot_images(
16
                images=batch["img"][:, 3:],
17
                labels=batched_preds,
18
                paths=batch["im_file"],
19
                fname=self.save_dir / f"val_batch_ir{ni}_pred.jpg",
20
                names=self.names,
21
                on_plot=self.on_plot,
22
            )