Network architecture different between paper and code

[weiguochow]

Hi， thank you very much for sharing your code. I am little confused that about the network depicted in the paper and code.

                um_in = tf.concat([hg_outs, hm_out, hm3_out], axis=-1)
                um_in = _residual(_residual(um_in, 256))

                um_in_mask = tf.concat([hg_outs, hm_out, hm3_out], axis=-1)
                mask = tf.tile(tf.less(tiny_dm, -0.9), (1,1,1,um_in_mask.get_shape()[-1].value))
                um_in_mask = tf.where(mask, tf.zeros_like(um_in_mask), um_in_mask)
                um_in_mask = _residual(_residual(um_in_mask, 256))

There is two consecutive residual model flowing um_in and um_in_mask, respectively. However, I cannot find in the figure 1 of the paper.

Thank you!

[melonwan]

Hey.
Thank you for your interest. Yes, the mask trick, together with we are also treating coordinate as input feature feed them to the network, are not mentioned in the paper due to the limitation of paper length.

[weiguochow]

Thank you for your response. I still have a question, that is, the image is cropped according to the root joint of the hand during training process. But for testing process, the root joint can not be obtained before fed to the network. Although we can use center of mass instead, but I find it will loss some accuracy.
Thank you vary much！!！

[melonwan]

for nyu dataset, we've trained another independent network for bounding box regression, but it is not part of the current project.