Probability Model-Adaptive Coding of Point Clouds with Octree Decomposition

彬彬

1 Introduction

An important observation is that when compressing point clouds

using Octree (OT) decomposition based compression algorithms

[Peng and Kuo 2005; Huang et al. 2008], the octree nodes with

only one non-empty child (single-point nodes), occur in an increasing

frequency as the cell subdivision goes deeper. The commonly

employed entropy codec uses a probability model which keeps updating

during the coding process. However, as illustrated in Fig.1

(b), the symbol distribution keeps varying and thus the probability

model trained online is seldom perfectly matched with the real statistical

distribution. Thus, there is still much room left to further

save the bitrates when using these codecs.

Figure 1: Our codec divides the octree into three portions and independently

compresses them with the most appropriate probability

model. The line chart in (a) is the octree symbol distribution of

Stanford’s bunny model. Its peak points correspond to the singlepoint

nodes which occur with an overwhelming probability. From

the line chart of (b), it is clear that our probability model (colorful

line) matches the distribution of single-point nodes (grey line) better

than the online trained probability model (black line) used by

the adaptive entropy codec

2 Algorithm and Experimental Results

Following [Huang et al. 2008], we use the 8-bit long occupancy

code to represent each cell subdivision, which uses a 1-bit flag to

signify whether a child cell is nonempty. Then all the occupancy

codes are included in symbol set S0 = {i}, i = 1...255. And

single-point nodes can be represented by symbol set S1 = {2i},

i = 0...7. As single-point nodes occur with an overwhelming probability

near the bottom of the octree, it is obvious that the octree

can be divided into upper and lower portions which are compressed

with S0 or S1 respectively. In the deeper layers of the upper part,

however, S1 symbols occur in a much higher probability than the

other S0 symbols. These layers cannot be compressed efficiently

with S0. Thus, we further divide the upper portion into 2 portions

and define a new symbol set S2 = {S1,X} for the deeper one.

The additional symbol X in S2 represents all the symbols that are

included in S0 and excluded from S1. The actual occupancy codes

are encoded in another pass with S0. Therefore, we divide the octree

into 3 portions, P0, P1 and P2, and compress them with S0,

S1 and S2, respectively.

晃晃

八方捷报，四海增辉。 一元复始，九州同庆； 八方和协，四季平安。 万里江山，增奇添彩； 四化图景，流翠飞红。

奇

此地無銀。。。

C.R.CAN

很经典，很实用，学习了!

奇

有意思！学习了！

奇

不错 非常经典  实用

C.R.CAN

俺是新人,这厢有礼了!

C.R.CAN

百度的叫度娘，网易的叫易娘，新浪内部还在为是叫新娘还是浪娘而争论不休！……不管你们是企鹅的额娘，豆瓣的伴娘，还是华为的伪娘，都要记得，淘宝才是你们的亲娘啊！亲！！

晃晃

无聊时可以刷屏幕 灌水 也可以试试 帖子的标题究竟可以写多长