The evaluation

[wangping1408]

Hello, about the evaluation in your paper, I wonder if you could give me the specific calculation process. I encountered some problems when calculating my result evaluation, thank you.

[TheSouthFrog]

Do you mean calculating the NME?
You can use this function to test NME normalized by inter-ocular distance:

def calculate_NME(self, logits, labels):
        logits = tf.cast(logits,tf.float32)
        labels = tf.cast(labels,tf.float32)
        subtract_square_distance = tf.square(tf.subtract(logits, labels))
        mean_distance = tf.reduce_mean([tf.sqrt(tf.add(subtract_square_distance[:, column],
                                                       subtract_square_distance[:, column+1])) for column in range(0, 136, 2)], axis=0)

        outer_eye_x = tf.square(tf.subtract(labels[:, 72], labels[:, 90]))
        outer_eye_y = tf.square(tf.subtract(labels[:, 73], labels[:, 91]))
        inter_ocular_distance = tf.sqrt(tf.add(outer_eye_x, outer_eye_y))
        normalized_mean_error = tf.divide(mean_distance, inter_ocular_distance,
                                          name='normalized_mean_error')
        loss_ = tf.reduce_mean(normalized_mean_error)

        return loss_