machine-learning - OpenPose 如何通过输出形状和真实不匹配来实现其损失函数？

Question

我最近一直在实现一个基于 OpenPose 的模型。在 OpenPose 中，它使用 VGG 作为其主干模型来提取特征图，但 VGG 包含最大池化层，这会将输出的形状减少到 1/4。下面是 OpenPose 的模型结构：

VGGOpenPose(
  (model0): OpenPose_Feature(
    (model): Sequential(
      (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (1): ReLU(inplace=True)
      (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (3): ReLU(inplace=True)
      (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
      (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (6): ReLU(inplace=True)
      (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (8): ReLU(inplace=True)
      (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
      (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (11): ReLU(inplace=True)
      (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (13): ReLU(inplace=True)
      (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (15): ReLU(inplace=True)
      (16): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (17): ReLU(inplace=True)
      (18): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
      (19): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (20): ReLU(inplace=True)
      (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (22): ReLU(inplace=True)
      (23): Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (24): ReLU(inplace=True)
      (25): Conv2d(256, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (26): ReLU(inplace=True)
    )
  )
  (model1_1): Sequential(
    (0): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
    (7): ReLU(inplace=True)
    (8): Conv2d(512, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model2_1): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model3_1): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model4_1): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model5_1): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model6_1): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 38, kernel_size=(1, 1), stride=(1, 1))
  )
  (model1_2): Sequential(
    (0): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1))
    (7): ReLU(inplace=True)
    (8): Conv2d(512, 19, kernel_size=(1, 1), stride=(1, 1))
  )
  (model2_2): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 19, kernel_size=(1, 1), stride=(1, 1))
  )
  (model3_2): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 19, kernel_size=(1, 1), stride=(1, 1))
  )
  (model4_2): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 19, kernel_size=(1, 1), stride=(1, 1))
  )
  (model5_2): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 19, kernel_size=(1, 1), stride=(1, 1))
  )
  (model6_2): Sequential(
    (0): Conv2d(185, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (1): ReLU(inplace=True)
    (2): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (3): ReLU(inplace=True)
    (4): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (5): ReLU(inplace=True)
    (6): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (7): ReLU(inplace=True)
    (8): Conv2d(128, 128, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3))
    (9): ReLU(inplace=True)
    (10): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))
    (11): ReLU(inplace=True)
    (12): Conv2d(128, 19, kernel_size=(1, 1), stride=(1, 1))
  )
)

在原始论文中，它说groundtruth热图和paf与输入图像具有相同的宽度和高度。 OpenPose：使用部分亲和场的实时多人 2D 姿势估计

我已经在 Python 中搜索了一些 OpenPose 的实现。他们中的大多数使用 element-wise loss function 来计算输出和 groundtruth label 之间的损失，就像论文中提到的函数一样：

openpose中的损失函数

我想知道 OpenPose 的输出是否与输入图像的大小不同，以及 OpenPose 是如何计算输出和 groundtruth heatmap/paf 之间的损失函数的？