1

我使用 FDDB 人脸图片来训练 mtcnn 进行人脸检测。在 pnet 中，对于边界框回归，acc 保持在 60% 左右。有什么不对？

时期 397/400 1200/1200 [===============================] - 0s 131us/样本 - 损失：0.5068 - conv4_1_loss : 5.4316e-06 - conv4_2_loss: 0.5052 - conv4_1_accuracy: 1.0000 - conv4_2_accuracy: 0.5850 Epoch 398/400 1200/1200 [========================= =====] - 0s 131us/样本 - 损失：0.4350 - conv4_1_loss：3.8598e-06 - conv4_2_loss：0.4358 - conv4_1_accuracy：1.0000 - conv4_2_accuracy：0.5950 Epoch 399/400 1200/1200 [====== =====================] - 0s 131us/样本 - 损失：0.8905 - conv4_1_loss：5.0222e-06 - conv4_2_loss：0.8863 - conv4_1_accuracy：1.0000 - conv4_2_accuracy： 0.5817 纪元 400/400 1200/1200 [==============================] - 0s 124us/样本 - 损失：1.8505 - conv4_1_loss：3.0373e-04 - conv4_2_loss：1.8358 - conv4_1_accuracy：1.0000 - conv4_2_accuracy：0.5817

class P_Net(keras.Model):
    def __init__(self):
        super(P_Net, self).__init__(name="P_Net")
        # Define layers here.
        self.conv1 = keras.layers.Conv2D(10, (3, 3), name="conv1")
        self.prelu1 = keras.layers.PReLU(tf.constant_initializer(0.25), shared_axes=[1, 2], name="prelu1")
        self.pool1 = keras.layers.MaxPooling2D((2, 2), name="pool1")
        self.conv2 = keras.layers.Conv2D(16, (3, 3), name="conv2")
        self.prelu2 = keras.layers.PReLU(tf.constant_initializer(0.25), shared_axes=[1, 2], name="prelu2")
        self.conv3 = keras.layers.Conv2D(32, (3, 3), name="conv3")
        self.prelu3 = keras.layers.PReLU(tf.constant_initializer(0.25), shared_axes=[1, 2], name="prelu3")
        self.cls_output = keras.layers.Conv2D(2, (1, 1), activation="softmax", name="conv4_1")
        self.bbox_pred = keras.layers.Conv2D(4, (1, 1), name="conv4_2")
        #self.landmark_pred = keras.layers.Conv2D(10, (1, 1), name="conv4_3")

    def call(self, inputs):
        # Define your forward pass here,
        # using layers you previously defined (in `__init__`).
        x = self.conv1(inputs)
        x = self.prelu1(x)
        x = self.pool1(x)
        x = self.conv2(x)
        x = self.prelu2(x)
        x = self.conv3(x)
        x = self.prelu3(x)
        return [self.cls_output(x), self.bbox_pred(x)]#, self.landmark_pred(x)]

    def get_summary(self, input_shape): 
        inputs = keras.Input(input_shape)
        model = keras.Model(inputs, self.call(inputs))
        return model

对于边界框 lebal，在 Positive train_set 中，我使用 FDDB 数据提供的可标记 x1,y1 x2,y2，只需根据图片的比例调整它们的大小。这是错的吗？Negative train_set，我设置 box lebal [0,0,0,0]

像这样：['./pos/20020816bigimg_932.jpg', 1, ['0.2857142857142857', '2.2857142857142856', '12.285714285714286', '14.285714285714286']]

4

0 回答 0