Question

我将通过考虑它们的时间相关性来完善与视频帧相关的顺序热图。我使用了ConvLSTM2D，但是输出看起来像输入平均值！精度小于输入。每个热图显示每个帧中每个关节的空间概率。你能帮我找到问题吗？

ModelInputSize = 300
TimeStep = 7
Joint= 16
def get_model(): 

X = Input(shape=(TimeStep, ModelInputSize, ModelInputSize, Joint),
                 batch_shape=(BatchSize,TimeStep, ModelInputSize, ModelInputSize, Joint),name='X')



temp = ConvLSTM2D(filters=Joint, kernel_size=(3, 3),  # 16
           padding='same', return_sequences=True)(X)



temp = BatchNormalization()(temp)

Y = Conv3D(filters=Joint, kernel_size=(3, 3, 3),  # 16
                 activation='sigmoid',  # 'sigmoid',
                 padding='same', data_format='channels_last')(temp)

Y = Conv3D(filters=Joint, kernel_size=(1, 1, 1),  # 16
           activation='sigmoid',  # 'sigmoid',
           padding='same', data_format='channels_last')(Y)

temp = Conv3D(filters=Joint, kernel_size=(1,1,1),  # 16
           activation='sigmoid',  # 'sigmoid',
           padding='same', data_format='channels_last')(X)

Y = keras.layers.Add()([temp, Y])

model = Model(input=X, output=Y)
model.compile(loss='mean_squared_error', optimizer='sgd')  # 'sgd'
model.summary()
return model

使用ConvLstm2D优化热图

0 个答案: