Question about similarities computation

[hi-zhengcheng]

Thanks for your work, it gives me a good example on how to fuse local features. When reading through the source code, I just find following places where I can't understand clearly, hope you can help me out.

Code 1

ELGS/utils/pointnet_gab_util.py

Line 91 in aae35e5

alpha = w_1 + tf.transpose(w_2, [0, 2, 1])

When computing the similarities of neighbor points in local area, I think dot product is the right way just as the formula(3) in the paper. But here, it uses a plus operation instead. So would you please share some ideas or experience about why plus is used here ?

Code 2

ELGS/models/sem_seg_model.py

Line 140 in aae35e5

energy_new=tf.tile(D, multiples=[1, 1,energy.shape[2]])-energy

I didn't get the idea why this subtraction is used here to compute the energy_new. Could you give some explanation on it ?

Thanks again for your nice work.