CHOG算法提出原始论文
需积分: 15 156 浏览量
2016-01-27
01:35:37
上传
评论
收藏 699KB PDF 举报
CHoG: Compressed Histogram of Gradients
A Low Bit-Rate Feature Descriptor
Vijay Chandrasekhar
†
Gabriel Takacs
†
David Chen
†
Sam Tsai
†
Radek Grzeszczuk
‡
Bernd Girod
†
†
Stanford University
Information Systems Lab
{vijayc,gtakacs,dmchen,sstsai,bgirod}@stanford.edu
‡
Nokia Research Center
Palo Alto, CA
radek.grzeszczuk@nokia.com
Abstract
Establishing visual correspondences is an essential com-
ponent of many computer vision problems, and is often done
with robust, local feature-descriptors. Transmission and
storage of these descriptors are of critical importance in
the context of mobile distributed camera networks and large
indexing problems. We propose a framework for comput-
ing low bit-rate feature descriptors with a 20× reduction in
bit rate. The framework is low complexity and has signifi-
cant speed-up in the matching stage. We represent gradient
histograms as tree structures which can be efficiently com-
pressed. We show how to efficently compute distances be-
tween descriptors in their compressed representation elimi-
nating the need for decoding. We perform a comprehensive
performance comparison with SIFT, SURF, and other low
bit-rate descriptors and show that our proposed CHoG de-
scriptor outperforms existing schemes.
1. Introduction
Local image features have become pervasive in the areas
of computer vision and image retrieval. These features are
increasingly finding applications in real-time object recog-
nition [1], 3D reconstruction [2], panorama stitching [3],
robotic mapping [4], and video tracking [5]. Depending
on the application, either transmission or storage issues (or
both) can limit the speed of computation or the size of
databases. In the context of mobile devices (camera phones)
or distributed camera networks, communication and power
costs are significant for transmitting information between
nodes. Feature compression is hence vital for reduction in
storage, latency and transmission.
Server-side Storage: Image retrieval applications need
query images to be matched against databases of millions of
features stored at application servers. Feature compression
can yield significant savings in storage space,
Application Latency: When data are sent over a net-
work, the system latency can be reduced by sending fewer
bits resulting from compression of image features.
Data Transmission: For mobile applications, band-
width is a limiting factor. Feature compression can reduce
the amount of data transmitted over wireless channels and
backhaul links in a mobile network.
Motivated by these demands, we propose a new frame-
work for computing low bit-rate feature descriptors.
1.1. Prior Work
Research on robust local descriptors continues to be a
very active area of computer vision. Of the many proposed
descriptors, SIFT [6] is probably the most commonly used.
Other popular descriptors include GLOH by Mikolajczyk
and Schmid [7], and SURF by Bay et al. [8]. The review
paper by Mikolajczyk et al. [9] provides a comprehensive
analysis of several descriptors. Winder and Brown [10]
also investigate various published descriptors, and propose
a framework for optimizing parameters in the descriptor
computation process.
Low-bit-rate descriptors are of increasing interest in the
vision community. Often, feature data are reduced by de-
creasing the dimensionality of descriptors. Ke and Suk-
thankar [11] investigate dimensionality reduction via Prin-
ciple Component Analysis. Hua et al. [12] propose a
scheme that uses Linear Discriminant Analysis. Takacs et
al. [1] quantize and entropy code SURF feature descriptors
for reducing their bit rate. Chandrasekhar et al. [13] pro-
pose a general framework for transform coding image fea-
tures. Chuohao et al. [14] reduce the bit rate of descriptors
by using random projections on SIFT descriptors to build
binary hashes. Descriptors are then compared using their
binary hashes. Shakhnarovich, in his thesis [15], uses a ma-
chine learning technique called Similarity Sensitive Coding
to train binary codes on image patches.
1.2. Contributions
We first show how to capture gradient statistics from
canonical patches in a histogram. In contrast to prior work,
we explicity exploit the underlying gradient statistics that
1
评论0
最新资源