• OctAttention
• • our contributions
  • file structure
  • environment
  • command
  • performance
  • Cite from：
  • contributors

OctAttention

key words: octree based, lossless, deeplearning, point cloud compression

OctAttention, is a kind of point cloud lossless compression method, based on octree. It encodes octree symbol sequences in a way by gathering the information of sibling and ancestor nodes. Compared to the previous state-of-the-art works, the approach obtains a 10%-35% BD-Rate gain on object point cloud dataset (e.g. MPEG 8i, MVUB), and saves 95% coding time compared to the voxel-based baseline.

our contributions

1.transplant from pytorch to tensorflow
2.test with several PC files, on pytorch and tensorflow
3.benchmark test

file structure

root
└── OctAttention-lidar.zip: pytorch code, lidar version
└── OctAttention-obj.zip: pytorch code, object version
└── TensorFlow: tensorflow code, models included
└── OctAttention Octree-based Large-scale Contexts Model for Point Cloud Compression.pdf: origional paper
└── datasets: see the train1.zip in dataset part

environment

1. pytorch
   refer to OctAttention-obj-pytorch/README.md

2. tensorflow
   refer to requirements.txt

command

• tensorflow:

cd TensorFlow

training:

python octAttention.py

encode:

python encoder.py --input="/userhome/PCGCv1/pytorch_eval/28_airplane_0270.ply" --ckpt_dir="checkpoint_TF_1024"

decode:

python decoder.py --input="/userhome/PCGCv1/pytorch_eval/28_airplane_0270.ply" --ckpt_dir="checkpoint_TF_1024"

performance

• test on some PC files (you can get them from here). From the result below, we can see that, encoding in pytorch is faster than tensorflow, maybe because of the fast operation of linear module of pytorch. And bpip of tensorflow is closed to pytorch.

PC files	TF_bpip	TF_enc_time	TF_dec_time	PT_enc_time	PT_dec_time	PT_bpip
redandblack_vox10_1550	0.779	55.124	6516.339	17.245	2147.175	0.737
longdress_vox10_1300	0.697	62.204	7346.297	18.843	2394.428	0.665
basketball_player_vox11_00000200	0.662	211.927	24562.372	55.532	8162.619	0.631
loot_vox10_1200	0.658	57.832	6874.465	17.22	2213.474	0.628
dancer_vox11_00000001	0.645	185.862	21693.571	48.242	6993.841	0.613
soldier_vox10_0690	0.702	78.737	9387.194	23.492	3048.636	0.67

• benchmark test on pytorch below for more details. From the result, we can see that for dense PCs, the method can achieve excellent compression rate, while for sparse PCs, the performance is not so good. The encoding time is short, but the decoding time is long.

file	binsize(b)	bit per oct	bpip	enc_time	oct len	dec_time
sarah_vox10_0023	974264	2.31	0.73	1.32	422648	3150.8
sarah_vox9_0023	224240	2.28	0.75	0.23	98425	734.5
phil_vox10_0139	1268184	2.51	0.8	1.16	504674	3800.3
phil_vox9_0139	298848	2.56	0.84	0.37	116745	873
Egyptian_mask_vox12	4085240	3.97	14.98	2.4	1027894	7693.1
Staue_Klimt_vox12	5734520	4.34	11.48	3.93	1320263	9872.9
vox10_002719	156936	4.91	3.46	0.18	31943	239
redandblack_vox10_1550	558672	1.97	0.74	0.67	284301	2147.2
queen_vox10_0200	671832	2.07	0.67	0.76	324707	2433.8
soldier_vox10_0690	730048	1.79	0.67	1.04	408418	3048.6
longdress_vox10_1300	570616	1.78	0.67	0.75	320608	2394.4
basketball_player_vox11_00000200	1846088	1.72	0.63	2.47	1072585	8162.6
loot_vox10_1200	505848	1.69	0.63	0.69	300192	2213.5
dancer_vox11_00000001	1588384	1.68	0.61	2.18	945988	6993.8
vox10_000001	185368	4.39	3.86	0.12	42190	311.4
vox10_002215	132904	4.25	3.81	0.08	31252	228.5
vox10_001406	162136	4.7	3.56	0.11	34528	255.8
vox10_000832	189432	4.7	4.15	0.1	40335	298.8
vox10_000355	142232	4.54	3.15	0.13	31299	230.9

Cite from：

@inproceedings{OctAttention,
title={OctAttention: Octree-Based Large-Scale Contexts Model for Point Cloud Compression},
author={Fu, Chunyang and Li, Ge and Song, Rui and Gao, Wei and Liu, Shan},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
year={2022}
}

contributors

name: Ye Hua
email: yeh@pcl.ac.cn