UPDATE_OPS

[Theophile88]

The code isn't missing something like:

update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = tf.train.AdamOptimizer(learning_rate = 1e-5).minimize(total_loss,global_step = global_step)

to ensure correct statistics during BN at test time?

[Joker316701882]

@Theophile88
Yeah that's right!
Here the training_batch size is set as 1, so this can be considered as using local statistics instead of moving statistics. At testing time, BN is still in training mode, so local statistics are also applied.
I tried using moving statistics of BN with batch size >1, there is no big difference according to experiment.
And as other issues mentioned, one main problem for not good performance is that I should use 'original size image inference' at test time instead of resized image.