CV党福音:YOLOv8实现语义分割

前面我们得知YOLOv8不但可以实现目标检测任务,还包揽了分类、分割、姿态估计等计算机视觉任务。在上一篇博文中,博主已经介绍了YOLOv8如何实现分类,在这篇博文里,博主将介绍其如何将语义分割给收入囊中。

YOLOv8语义分割架构图

如下图所示,YOLOv8采用了一种分割头与检测头相结合的方式来进行语义分割,在这个过程中,其会输出目标检测框与语义分割蒙版。

在这里插入图片描述

输出结果图像如下:

在这里插入图片描述

经典分割模型结构

为了让我们更好的理解语义分割模型,我们以最经典的语义分割模型UNet为例,可以看到其最终的结果要与原图像大小相同,但最终的维度(n)会有差别,这与我们确定使用的mask的数量有关。
在这里插入图片描述

YOLOv8语义分割模型结构

语义分割模型的结构与YOLOv8的目标检测模型结构即为接近,区别在于最后的目标检测头替换为语义分割头,同时其最终的语义分割头也是具有三种尺度的:

[[15, 18, 21], 1, Segment, [nc, 32, 256]] # Segment(P3, P4, P5)

下图中对各个模块进行了编号,大家可以与yaml的模型文件进行对照
在这里插入图片描述

# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8-seg instance segmentation model. For Usage examples see https://docs.ultralytics.com/tasks/segment

# Parameters
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-seg.yaml' will call yolov8-seg.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]
  s: [0.33, 0.50, 1024]
  m: [0.67, 0.75, 768]
  l: [1.00, 1.00, 512]
  x: [1.00, 1.25, 512]

# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  - [-1, 3, C2f, [128, True]]
  - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  - [-1, 6, C2f, [256, True]]
  - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  - [-1, 6, C2f, [512, True]]
  - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  - [-1, 3, C2f, [1024, True]]
  - [-1, 1, SPPF, [1024, 5]] # 9

# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 6], 1, Concat, [1]] # cat backbone P4
  - [-1, 3, C2f, [512]] # 12

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [[-1, 4], 1, Concat, [1]] # cat backbone P3
  - [-1, 3, C2f, [256]] # 15 (P3/8-small)

  - [-1, 1, Conv, [256, 3, 2]]
  - [[-1, 12], 1, Concat, [1]] # cat head P4
  - [-1, 3, C2f, [512]] # 18 (P4/16-medium)

  - [-1, 1, Conv, [512, 3, 2]]
  - [[-1, 9], 1, Concat, [1]] # cat head P5
  - [-1, 3, C2f, [1024]] # 21 (P5/32-large)

  - [[15, 18, 21], 1, Segment, [nc, 32, 256]] # Segment(P3, P4, P5)

YOLOv8检测头(可忽略)

那么我们看下这个分割头到底是如何定义的,分割头继承了检测头:

检测头代码如下:我们可以看到ultralytics更新了检测头(加入了YOLOv10,博主这里将该方法删掉了,因为用不到),其创新点为混合匹配机制,故在检测头中多出了forward_end2end

class Detect(nn.Module):
    """YOLOv8 Detect head for detection models."""

    dynamic = False  # force grid reconstruction
    export = False  # export mode
    end2end = False  # end2end
    max_det = 300  # max_det
    shape = None
    anchors = torch.empty(0)  # init
    strides = torch.empty(0)  # init

    def __init__(self, nc=80, ch=()):
        """Initializes the YOLOv8 detection layer with specified number of classes and channels."""
        super().__init__()
        self.nc = nc  # number of classes
        self.nl = len(ch)  # number of detection layers
        self.reg_max = 16  # DFL channels (ch[0] // 16 to scale 4/8/12/16/20 for n/s/m/l/x)
        self.no = nc + self.reg_max * 4  # number of outputs per anchor
        self.stride = torch.zeros(self.nl)  # strides computed during build
        c2, c3 = max((16, ch[0] // 4, self.reg_max * 4)), max(ch[0], min(self.nc, 100))  # channels
        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(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:
            self.one2one_cv2 = copy.deepcopy(self.cv2)
            self.one2one_cv3 = copy.deepcopy(self.cv3)

    def forward(self, x):
        """Concatenates and returns predicted bounding boxes and class probabilities."""
        if self.end2end:
            return self.forward_end2end(x)

        for i in range(self.nl):
            x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
        if self.training:  # Training path
            return x
        y = self._inference(x)
        return y if self.export else (y, x)


        y = self._inference(one2one)
        y = self.postprocess(y.permute(0, 2, 1), self.max_det, self.nc)
        return y if self.export else (y, {"one2many": x, "one2one": one2one})

    def _inference(self, x):
        """Decode predicted bounding boxes and class probabilities based on multiple-level feature maps."""
        # Inference path
        shape = x[0].shape  # BCHW
        x_cat = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2)
        if self.dynamic or self.shape != shape:
            self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
            self.shape = shape

        if self.export and self.format in {"saved_model", "pb", "tflite", "edgetpu", "tfjs"}:  # avoid TF FlexSplitV ops
            box = x_cat[:, : self.reg_max * 4]
            cls = x_cat[:, self.reg_max * 4 :]
        else:
            box, cls = x_cat.split((self.reg_max * 4, self.nc), 1)

        if self.export and self.format in {"tflite", "edgetpu"}:
            # Precompute normalization factor to increase numerical stability
            # See https://github.com/ultralytics/ultralytics/issues/7371
            grid_h = shape[2]
            grid_w = shape[3]
            grid_size = torch.tensor([grid_w, grid_h, grid_w, grid_h], device=box.device).reshape(1, 4, 1)
            norm = self.strides / (self.stride[0] * grid_size)
            dbox = self.decode_bboxes(self.dfl(box) * norm, self.anchors.unsqueeze(0) * norm[:, :2])
        else:
            dbox = self.decode_bboxes(self.dfl(box), self.anchors.unsqueeze(0)) * self.strides

        return torch.cat((dbox, cls.sigmoid()), 1)

YOLOv8分割头

分割头代码如下:

class Segment(Detect):
    """YOLOv8 Segment head for segmentation models."""

    def __init__(self, nc=80, nm=32, npr=256, ch=()):
        """Initialize the YOLO model attributes such as the number of masks, prototypes, and the convolution layers."""
        super().__init__(nc, ch)
        self.nm = nm  # number of masks
        self.npr = npr  # number of protos
        self.proto = Proto(ch[0], self.npr, self.nm)  # protos

        c4 = max(ch[0] // 4, self.nm)
        self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)

    def forward(self, x):
        """Return model outputs and mask coefficients if training, otherwise return outputs and mask coefficients."""
        p = self.proto(x[0])  # mask protos
        bs = p.shape[0]  # batch size

        mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
        x = Detect.forward(self, x)
        if self.training:
            return x, mc, p
        return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))

输入到分割头的图像存储在list中,共有三个不同尺度,这与YOLOv8目标检测是相同的

在这里插入图片描述

上述第一个操作便是Proto操作 ,传入的是第一尺度的输出特征图,Proto的功能是针对x[0]进行卷积,将原来80x80大小的feature通过上采样变为160x160,这个图像是基础蒙版(mask)。

p = self.proto(x[0])
Proto(
  (cv1): Conv(
    (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (act): SiLU(inplace=True)
  )
  (upsample): ConvTranspose2d(64, 64, kernel_size=(2, 2), stride=(2, 2))
  (cv2): Conv(
    (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (act): SiLU(inplace=True)
  )
  (cv3): Conv(
    (conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1))
    (act): SiLU(inplace=True)
  )
)

随后将x输入到cv4模块(期内包含3个模块组成))(即图像中的三个不同尺度的操作),cv4结构如下:

ModuleList(
  (0): Sequential(
    (0): Conv(
      (conv): Conv2d(64, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))   #CBS模块
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))   #CBS模块
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1))     #用于通道维度转换,Conv2d
  )
  (1): Sequential(
    (0): Conv(
      (conv): Conv2d(128, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1))
  )
  (2): Sequential(
    (0): Conv(
      (conv): Conv2d(256, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1))
  )
)

其过程如下图所示:
在这里插入图片描述
得到的即为mask
随后进入检测的前向传播过程,因为YOLOv8本身就是做的检测,因此这个结果还是要进入检测头:

for i in range(self.nl):
    x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)

cv2中包含三个模块,最终的输出大小不变,通道数均为64,即为(64,80,80)(64,40,40)(64,20,20)

ModuleList(
  (0): Sequential(
    (0): Conv(
      (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
  )
  (1): Sequential(
    (0): Conv(
      (conv): Conv2d(128, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
  )
  (2): Sequential(
    (0): Conv(
      (conv): Conv2d(256, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))
  )
)

cv3中也是三个模块,图像大小依旧不变,通道维度变为80,即(80,80,80)(80,40,40)(80,20,20)

ModuleList(
  (0): Sequential(
    (0): Conv(
      (conv): Conv2d(64, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(80, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(80, 80, kernel_size=(1, 1), stride=(1, 1))
  )
  (1): Sequential(
    (0): Conv(
      (conv): Conv2d(128, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(80, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(80, 80, kernel_size=(1, 1), stride=(1, 1))
  )
  (2): Sequential(
    (0): Conv(
      (conv): Conv2d(256, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (1): Conv(
      (conv): Conv2d(80, 80, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (act): SiLU(inplace=True)
    )
    (2): Conv2d(80, 80, kernel_size=(1, 1), stride=(1, 1))
  )
)

最终将其使用torch.cat进行拼接,得到(144,80,80)(144,40,40)(144,20,20)

随后便是推理的后处理过程,即对输出的这三个尺度的图像进行解码:

self.no = nc + self.reg_max * 4,其中reg_max是根据YOLOv8不同模型大小设定的,即 scale 4/8/12/16/20 for n/s/m/l/x),此处reg_max=16

def _inference(self, x):
        """Decode predicted bounding boxes and class probabilities based on multiple-level feature maps."""
        # Inference path
        shape = x[0].shape  # BCHW  (144,80,80)
        x_cat = torch.cat([xi.view(shape[0], self.no, -1) for xi in x], 2)#(1,144,8400)8400=80*80+40*40+20*20
        if self.dynamic or self.shape != shape:
            self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors(x, self.stride, 0.5))
            self.shape = shape  #

        if self.export and self.format in {"saved_model", "pb", "tflite", "edgetpu", "tfjs"}:  # avoid TF FlexSplitV ops#不执行
            box = x_cat[:, : self.reg_max * 4]
            cls = x_cat[:, self.reg_max * 4 :]
        else:
            box, cls = x_cat.split((self.reg_max * 4, self.nc), 1)

        if self.export and self.format in {"tflite", "edgetpu"}:#不执行
            # Precompute normalization factor to increase numerical stability
            # See https://github.com/ultralytics/ultralytics/issues/7371
            grid_h = shape[2]
            grid_w = shape[3]
            grid_size = torch.tensor([grid_w, grid_h, grid_w, grid_h], device=box.device).reshape(1, 4, 1)
            norm = self.strides / (self.stride[0] * grid_size)
            dbox = self.decode_bboxes(self.dfl(box) * norm, self.anchors.unsqueeze(0) * norm[:, :2])
        else:
            dbox = self.decode_bboxes(self.dfl(box), self.anchors.unsqueeze(0)) * self.strides

        return torch.cat((dbox, cls.sigmoid()), 1)

self.anchors为torch.Size([2, 8400]), self.strides为torch.Size([1, 8400])
shapetorch.Size([1, 144, 80, 80]) 144=64+80,这个64是预测的box的值,最后还要进行转换

根据x_cat进行拆分,得到预测的boxclsbox即为(1,64,8400)cls(1,80,8400)
随后通过DEL模块对box进行分解:

dbox = self.decode_bboxes(self.dfl(box), self.anchors.unsqueeze(0)) * self.strides

得到的box即为(1,4,8400)

在这里插入图片描述
DEL中的Conv2d没有梯度,即参数不会更新,这个模块作用便是将64分解为4*16,进而得到4*1
最后将 dboxcls(类别)返回

return torch.cat((dbox, cls.sigmoid()), 1)

其维度为(84,8400)84=80+48400代表预测的目标个数

最终返回数据:

return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))

其中mc为(1,32,8400)x是一个元组,x[0]为(1,84,8400)x[1】为列表,包含(1,144,80,80)(1,144,40,40),(1,144,20,20),p为(1,32,160,160)p为基础蒙版

返回的数据:

在这里插入图片描述

其中(1,32,8400)即为预测的mask

非极大值抑制

在推理过程中,博主使用的图像大小为(3,480,640),所有最后得到的数据维度为(1,116,6300)其中,116=84(80+4)+32,这是因为YOLOv8中不仅要完成语义分割还要实现目标检测,其中(1,32,6300)是用于语义分割的。
6300=60*80+30*40+15*20

下面的分解代码证实了这一点,即mask的数量(nm)为116-80-4=32mask开始坐标为80+4=84

nm = prediction.shape[1] - nc - 4  # number of masks
mi = 4 + nc

设定输出:

output = [torch.zeros((0, 6 + nm), device=prediction.device)] * bs

得到的output为38个值,其中38=4+class_score+class+32(保存检测与分割结果)
随后筛选出的大于阈值的类别,得到36个,即(1,36,116),这里的36指的是符合的个数,是从6300中筛选出的。

box,类别 以及分割mask分开:

box, cls, mask = x.split((4, nc, nm), 1)

box(36,4)cls(36,80)mask(36,32)

随后再从类别中选出最大的

conf, j = cls.max(1, keepdim=True)
x = torch.cat((box, conf, j.float(), mask), 1)[conf.view(-1) > conf_thres]

得到的conf为分值,j为坐标(代表类别),维度均为(36,1),并将这些数据再次拼接到一起,得到(36,38),其中36为目标个数,384+1+1+32,即 box+conf+cls_id+mask

c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes,max_wh是定义的,值为7860
scores = x[:, 4]  # scores
if rotated:
      boxes = torch.cat((x[:, :2] + c, x[:, 2:4], x[:, -1:]), dim=-1)  # xywhr
      i = nms_rotated(boxes, scores, iou_thres)
else:#执行这个,对box进行非极大值抑制,这个是调用了torch的包
      boxes = x[:, :4] + c  # boxes (offset by class)
      i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
i = i[:max_det]#max_det=300,即最多只能预测300个目标

output[xi] = x[i]

返回的 itensor([20, 24, 3, 32, 34], device='cuda:0'),这里给出的i36个中经过筛选后的检测框编号,最终将x中的目标筛选出存储到output中,可以看到output是一个列表,存放的是每个batch的结果,由于在预测时只输入一张图像,故里面只有一个数据,筛选出的结果为(5,38),即有5个目标。

在这里插入图片描述

后处理过程

那么,这个mask要如何使用呢。我们接下来看一下其后处理过程

def postprocess(self, preds, img, orig_imgs):

后处理过程中传入的参数为preds即预测的结果,即YOLOv8分割头输出的结果

在这里插入图片描述

img是输入的图像(归一化后的),orig_img是原始图像

在这里插入图片描述

在后处理过程的刚开始,便是利用非极大值抑制来筛选出部分数据:

p = ops.non_max_suppression(
            preds[0],
            self.args.conf,
            self.args.iou,
            agnostic=self.args.agnostic_nms,
            max_det=self.args.max_det,
            nc=len(self.model.names),
            classes=self.args.classes,
        )

得到的结果 p 即为(5,38)

判断preds[1]是否是tuple类型,是,则为preds[1][-1],即为(1,32,120,160)

proto = preds[1][-1] if isinstance(preds[1], tuple) else preds[1]

随后进行下面的循环(预测只有一张图像,故只有一轮)

for i, (pred, orig_img, img_path) in enumerate(zip(p, orig_imgs, self.batch[0])):
            if not len(pred):  # save empty boxes
                masks = None
            elif self.args.retina_masks:
                pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], orig_img.shape)
                masks = ops.process_mask_native(proto[i], pred[:, 6:], pred[:, :4], orig_img.shape[:2])  # HWC
            else:#执行的是这个分支
                masks = ops.process_mask(proto[i], pred[:, 6:], pred[:, :4], img.shape[2:], upsample=True)  # HWC
                pred[:, :4] = ops.scale_boxes(img.shape[2:], pred[:, :4], orig_img.shape)
            results.append(Results(orig_img, path=img_path, names=self.model.names, boxes=pred[:, :6], masks=masks))
        return results

我们可以看到mask的处理结果:

masks = ops.process_mask(proto[i], pred[:, 6:], pred[:, :4], img.shape[2:], upsample=True)

process_mask方法是如何处理的呢?
我们先看一下其传入的参数,proto(1,32,120,160)pred(5,38),取从638,即只是mask32维数据,即为(5,32),同时还有bbox(5,4)img.shape[2:]为宽高

def process_mask(protos, masks_in, bboxes, shape, upsample=False):
    c, mh, mw = protos.shape  # CHW 32 120 160
    ih, iw = shape #480,640
    masks = (masks_in @ protos.float().view(c, -1)).view(-1, mh, mw)  # CHW  (5,120,160)
    width_ratio = mw / iw #0.25
    height_ratio = mh / ih#0.25

    downsampled_bboxes = bboxes.clone()#克隆
    downsampled_bboxes[:, 0] *= width_ratio #对其进行缩放
    downsampled_bboxes[:, 2] *= width_ratio
    downsampled_bboxes[:, 3] *= height_ratio
    downsampled_bboxes[:, 1] *= height_ratio

    masks = crop_mask(masks, downsampled_bboxes)  # CHW
    if upsample:
        masks = F.interpolate(masks[None], shape, mode="bilinear", align_corners=False)[0]  # CHW
    return masks.gt_(0.0)
bboxs原始值:
tensor([[2.3550e+02, 1.1798e+02, 3.6113e+02, 3.4263e+02],
        [2.7596e-01, 1.5049e+02, 1.8605e+02, 4.1289e+02],
        [5.9170e+02, 1.7436e+02, 6.3966e+02, 3.0410e+02],
        [9.4319e+00, 1.5594e+02, 4.7882e+02, 4.7825e+02],
        [3.0707e+01, 1.3653e+02, 4.7746e+02, 4.7805e+02]], device='cuda:0')
        
bbox缩小后的值:
tensor([[5.8876e+01, 2.9494e+01, 9.0283e+01, 8.5657e+01],
        [6.8989e-02, 3.7624e+01, 4.6512e+01, 1.0322e+02],
        [1.4793e+02, 4.3590e+01, 1.5992e+02, 7.6024e+01],
        [2.3580e+00, 3.8984e+01, 1.1971e+02, 1.1956e+02],
        [7.6767e+00, 3.4133e+01, 1.1936e+02, 1.1951e+02]], device='cuda:0')

接下来便是将保证mask在Bbox内。

masks = crop_mask(masks, downsampled_bboxes)
def crop_mask(masks, boxes):
    _, h, w = masks.shape
    x1, y1, x2, y2 = torch.chunk(boxes[:, :, None], 4, 1)  # x1 shape(n,1,1)
    r = torch.arange(w, device=masks.device, dtype=x1.dtype)[None, None, :]  # rows shape(1,1,w)
    c = torch.arange(h, device=masks.device, dtype=x1.dtype)[None, :, None]  # cols shape(1,h,1)

    return masks * ((r >= x1) * (r < x2) * (c >= y1) * (c < y2))

得到的mask依旧为(5,120,160),随后对mask进行上采样,使其与原本的图像大小一样的,这里就已经是蒙版了,通过插值的方式进行上采样,得到的mask为(5,480,640)

mask的理解

在检测头(分割头)中输出的32 维的向量可以看作是与每个检测框关联的分割 mask 的系数或权重。

针对于分割头的输出 1x32x160x160,一个关键的概念是 prototype masks。它是一个固定数量(32)的基础mask,每个 mask 的尺寸为 160×160。这些基础 mask并不直接对应于任何特定的物体或类别,而是被设计为可以线性组合来表示任何可能的物体 mask

简单来说,模型不直接预测每个物体的完整 mask,而是预测一组基本的 masks(称为 prototype masks)以及每个物体如何组合这些 masks(权重/系数)。这种方法的好处是,模型只需要预测一个较小的 mask张量,然后可以通过简单的矩阵乘法将这些小 mask 组合成完整的物体 masks

大家可以把它类比于线性代数中基向量的概念,空间中的任何一个向量是不是都可以表示为一组基向量的线性组合,那么其中的 prototype masks32x160x160mask 张量可以把它理解为一组基向量,而之前在检测框中的 32维向量可以理解为组合这一组基向量的权重或者说系数。

当我们从检测头得到一个 32 维的向量,分割头得到 32 个基础 masks 时,这个 32 维的向量实际上表示了如何组合这些基础masks 来得到一个特定物体的 mask。具体来说,我们用这个 32 维向量对 32 个基础 masks进行线性组合,从而得到与检测框关联的最终 mask。简单来说,这就像你现在有 32 种不同的颜料,检测头给你一个配方(32 维向量),告诉你如何混合这些颜料来得到一个特定的颜色(最终的 mask)。

这样做的优点是我们不需要为每个检测框都预测一个完整的 mask,这个非常消耗内存和计算资源。相反,我们只需要预测一个相对较小的 32 维向量和一个固定数量的基础 masks,然后在后处理中进行组合即可。

结果可视化

最后附上将结果可视化的代码

import cv2
import numpy as np
from ultralytics import YOLO

def hsv2bgr(h, s, v):
    h_i = int(h * 6)
    f = h * 6 - h_i
    p = v * (1 - s)
    q = v * (1 - f * s)
    t = v * (1 - (1 - f) * s)

    r, g, b = 0, 0, 0

    if h_i == 0:
        r, g, b = v, t, p
    elif h_i == 1:
        r, g, b = q, v, p
    elif h_i == 2:
        r, g, b = p, v, t
    elif h_i == 3:
        r, g, b = p, q, v
    elif h_i == 4:
        r, g, b = t, p, v
    elif h_i == 5:
        r, g, b = v, p, q

    return int(b * 255), int(g * 255), int(r * 255)

def random_color(id):
    h_plane = (((id << 2) ^ 0x937151) % 100) / 100.0
    s_plane = (((id << 3) ^ 0x315793) % 100) / 100.0
    return hsv2bgr(h_plane, s_plane, 1)

if __name__ == "__main__":

    model = YOLO("yolov8n-seg.pt")

    img = cv2.imread("img.jpg")
    result = model(img)[0]
    names = result.names
    boxes = result.boxes.data.tolist()
    masks = result.masks

    h, w = img.shape[:2]

    for i, mask in enumerate(masks.data):

        mask = mask.cpu().numpy().astype(np.uint8)
        mask_resized = cv2.resize(mask, (w, h))

        label = int(boxes[i][5])
        color = np.array(random_color(label))

        colored_mask = (np.ones((h, w, 3)) * color).astype(np.uint8)
        masked_colored_mask = cv2.bitwise_and(colored_mask, colored_mask, mask=mask_resized)

        mask_indices = mask_resized == 1
        img[mask_indices] = (img[mask_indices] * 0.6 + masked_colored_mask[mask_indices] * 0.4).astype(np.uint8)


    for obj in boxes:
        left, top, right, bottom = int(obj[0]), int(obj[1]), int(obj[2]), int(obj[3])
        confidence = obj[4]
        label = int(obj[5])
        color = random_color(label)
        cv2.rectangle(img, (left, top), (right, bottom), color = color ,thickness=2, lineType=cv2.LINE_AA)
        caption = f"{names[label]} {confidence:.2f}"
        w, h = cv2.getTextSize(caption, 0, 1, 2)[0]
        cv2.rectangle(img, (left - 3, top - 33), (left + w + 10, top), color, -1)
        cv2.putText(img, caption, (left, top - 5), 0, 1, (0, 0, 0), 2, 16)

    cv2.imwrite("predict-seg.jpg", img)
    print("save done")

  • 8
    点赞
  • 19
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

彭祥.

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值