search

digitalbrain79 / person-reid

Tensorflow implementation of "An Improved Deep Learning Architecture for Person Re-Identification"
118 stars 61 forks source link

Does the tide convolution being performed correctly? #6

Open LDHo opened 6 years ago

LDHo commented 6 years ago

According to my understanding on the paper, in order for both images to be able to perform cross input neighborhood differences, the first two convolution layers must be tied in which their trained filters must be shared for two input images. Can I know whether the tide convolution is performed correctly as from my understanding, your codes are using totally different filters for both images.

stickOverCarrot commented 6 years ago

@LDHo I think we can replace [conv2_1 = tf.layers.conv2d(images2, 20, [5, 5], activation=tf.nn.relu, kernel_regularizer=tf.contrib.layers.l2_regularizer(weight_decay), name='conv2_1')] with [ conv2_1 = tf.layers.conv2d(images2, 20, [5, 5], activation=tf.nn.relu, kernel_regularizer=tf.contrib.layers.l2_regularizer(weight_decay), name='conv1_1',reuse=True)]. Other conv layers can share weight and bias by this way. In fact,we can try the both approaches and then compare the results to judge which is better!