search

hellochick / PSPNet-tensorflow

TensorFlow-based implementation of "Pyramid Scene Parsing Network".
326 stars 123 forks source link

eval_flip #20

Open FengLoveBella opened 6 years ago

FengLoveBella commented 6 years ago

Could you explain how the flipped_eval work? I put raw_output(image) and raw_output(flip_image) together, but I found I got a worse prediction than using only raw_output(image). The left is my code based on your flipped_eval.

    if flipped_aval:
        flipped_out = tf.image.flip_left_right(tf.squeeze(net2.outputs))
        flipped_out = tf.expand_dims(flipped_out, dim=0)
        raw_output = tf.add_n([raw_output, flipped_out])
        raw_output_up = tf.image.resize_bilinear(raw_output, tf.shape(self.image_batch)[1:3, ])
        raw_output_up = tf.argmax(raw_output_up, axis=3)
        pred = tf.expand_dims(raw_output_up, dim=3)
    else:
        raw_output = tf.image.resize_bilinear(raw_output, tf.shape(self.image_batch)[1:3,])
        raw_output = tf.argmax(raw_output, axis=3)
        pred = tf.expand_dims(raw_output, dim=3)