Contrastive Learning for 3D Point Clouds Classification and Shape Completion

. 2021 Nov 6;21(21):7392. doi: 10.3390/s21217392

Algorithm 1 Training Contrasitive AutoEncoder.

1:
for $e p o c h = 1, \dots t o t a l_e p o c h s$ do
2:
for $i, d a t a i n e n u m e r a t e (t r a i n_l o a d e r)$ do
3:
Extract online triplets and ground truth.
4:
$r_{a}$ , $r_{p}$ , $r_{n}$ = $t r a i n_l o a d e r [“ t r i p l e t^{”}]$
5:
$e_{a}$ , $e_{p}$ , $e_{n}$ = $E n c o d e r (r_{a}, r_{p}, r_{n})$
6:
$d e c_e_{a}$ , $d e c_e_{p}$ , $d e c_e_{n}$ = $D e c o d e r (e_{a}, e_{p}, e_{n})$
7:
Compute Chamfer Loss and triplet loss.
8:
$t r i p l e t_l o s s$ = $c o m p u t e_t r i p l e t_l o s s (e_{a}, e_{p}, e_{n})$
9:
$a n c h o r_c h$ = $c o m p u t e_c h a m f e r (r_a, d e c_e_{a})$
10:
$p o s_c h$ = $c o m p u t e_c h a m f e r (r_p, d e c_e_{p})$
11:
$n e g a t i v e_c h$ = $c o m p u t e_c h a m f e r (r_n, d e c_e_{n})$
12:
Calculate and backpropogate the total loss.
13:
$c h a m f e r_l o s s e s$ = $p o s i t i v e_c h$ + $n e g a t i v e_c h$ + $a n c h o r_c h$
14:
$t o t a l_l o s s$ = $t r i p l e t_l o s s$ + $c h a m f e r_l o s s e s$
15:
$t o t a l_l o s s . b a c k w a r d ()$
16:
end for
17:
end for