Why use tf.zeros_initializer() to init conv2d weights?

[GesilaA]

I have a puzzle about the tf.zeros_initializer() at line:85 in palm_detector.py conv_zeros = Conv2D(feature_n/2,2,strides=2,use_bias=False,kernel_initializer=tf.zeros_initializer())(tensor) Why this Conv2D operator should be initialized by tf.zeros_initializer()?