swook / GazeML

Gaze Estimation using Deep Learning, a Tensorflow-based framework.
MIT License
522 stars 140 forks source link

Is it a bug in elg.py when doing the merge? #65

Open ShukuiZhang opened 4 years ago

ShukuiZhang commented 4 years ago

This is great work but something doesn't make sense to me here. I copy paste the code as follows:

        x_next = x_now
        if do_merge:
            with tf.variable_scope('merge'):
                with tf.variable_scope('h'):
                    x_hmaps = self._apply_conv(h, self._hg_num_feature_maps, kernel_size=1, stride=1)
                with tf.variable_scope('x'):
                    x_now = self._apply_conv(x_now, self._hg_num_feature_maps, kernel_size=1, stride=1)
                x_next += x_prev + x_hmaps

x_now above is not used at all.
After reading the original Hourglass paper, I think the last line above should look like

                x_next = x_prev + x_hmaps + x_now