Table 2 Complexity and compression...

Table 2

Complexity and compression performance comparison among different methods on “longdress” point cloud (Anchor: G-PCC). The “L” denotes a lightweight transform network comprising of stacked convolutional layers, while “H” represents a heavyweight network incorporating self-attention layers.

Methods	Transform	#Param. ↓	GPU Mem. (GB) ↓	Enc. Time (s) ↓	Dec. Time (s) ↓	BD-PSNR (dB) ↑	BD-Rate (%) ↓
G-PCC	RAHT	-	-	0.700	0.509	-	-
G-PCC (TMC13v19)	RAHT	-	-	4.281	3.889	+2.12	−45.49
G-PCC (TMC13v23)	RAHT	-	-	5.062	4.315	+2.38	−49.75
Factorized	L	3.554M	1.903	0.148	0.153	+1.68	−37.22
Hyper	L	8.758M	1.911	0.210	0.192	+2.10	−44.33
Joint (Hyper + Autoregressive)	L	9.872M	1.848	0.189	64.079	+2.37	−53.17
Ours-Light w/o Global	L	9.872M	1.913	0.313	0.222	+2.17	−45.36
Ours-Light	L	18.103M	1.919	0.442	0.362	+2.57	−50.48
Ours	H	31.127M	2.491	0.513	0.391	+3.16	−57.90

Methods	Transform	#Param. ↓	GPU Mem. (GB) ↓	Enc. Time (s) ↓	Dec. Time (s) ↓	BD-PSNR (dB) ↑	BD-Rate (%) ↓
G-PCC	RAHT	-	-	0.700	0.509	-	-
G-PCC (TMC13v19)	RAHT	-	-	4.281	3.889	+2.12	−45.49
G-PCC (TMC13v23)	RAHT	-	-	5.062	4.315	+2.38	−49.75
Factorized	L	3.554M	1.903	0.148	0.153	+1.68	−37.22
Hyper	L	8.758M	1.911	0.210	0.192	+2.10	−44.33
Joint (Hyper + Autoregressive)	L	9.872M	1.848	0.189	64.079	+2.37	−53.17
Ours-Light w/o Global	L	9.872M	1.913	0.313	0.222	+2.17	−45.36
Ours-Light	L	18.103M	1.919	0.442	0.362	+2.57	−50.48
Ours	H	31.127M	2.491	0.513	0.391	+3.16	−57.90

[ViewLarge]