tristanguigue / universal-correspondence-network Goto Github PK

there is a piece of code in definition of the knn function

flat_feature2 = tf.reshape(feature2, (-1, feature_size_x * feature_size_y, feature_channels))
tf.subtract(tf.expand_dims(feature1_proj_at_corres, 2), tf.expand_dims(flat_feature2, 1)))

what does the code above do? I think that
the feature1_proj_at_corres should have the size of (Batch_size, nb_corr, channels)
and the flat_feature2 should have the size of (Batch_size, heigt* width,channels)
right?
so I don't know what do the tf.subtract() and the tf.expand_dims() do?
Looking forward to your reply. Thank you very much! :-)

The code in your layers.py is as follows(in the KNN definition part):
distance = tf.negative(
tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(
tf.expand_dims(feature1_proj_at_corres, 2), tf.expand_dims(flat_feature2, 1))), reduction_indices=3)))
flat_indices_nearest = tf.cast(tf.argmin(distance, 2), tf.int32)

We know the result of the tf.sqrt() is non-negative value, but you use a tf.negative() to make it become non-positive value. And then you use the tf.argmin() to choose the smallest value in dim_2.
But if numbers are non-positive, the smaller number has a bigger absolute value, which means the distance is bigger. So the result is not the most similar feature?
Looking forward to your reply~ Thank you very much :-)