![David C. Wyld et al. (Eds) : ITCS, CST, JSE, SIP, ARIA, DMS - 2015
pp. 25–31, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.50103
IMAGE RETRIEVAL USING VLAD WITH
MULTIPLE FEATURES
Pin-Syuan Huang, Jing-Yi Tsai, Yu-Fang Wang, and Chun-Yi Tsai
Department of Computer Science and Information Engineering,
National Taitung University, Taiwan, R.O.C.
{u10011104,u10011127,u10011139}@ms100.nttu.edu.tw;
cytsai@nttu.edu.tw
ABSTRACT
The objective of this paper is to propose a combinatorial encoding method based on VLAD to
facilitate the promotion of accuracy for large scale image retrieval. Unlike using a single
feature in VLAD, the proposed method applies multiple heterogeneous types of features, such as
SIFT, SURF, DAISY, and HOG, to form an integrated encoding vector for an image
representation. The experimental results show that combining complementary types of features
and increasing codebook size yield high precision for retrieval.
KEYWORDS
VLAD, SIFT, SURF, DAISY, HOG
1. INTRODUCTION
Image retrieval is one of the classical problems in computer vision and machine intelligence. The
challenge of image retrieval is mainly aimed at trade-off between computing costs and the
precision of retrieval due to the large scale data, including the large size of a single image and the
large number of all images. In this paper, we propose an combinatorial encoding algorithm based
on VLAD[6,9] with multiple features to achieve the goal of large scale image retrieval. VLAD
cluster all features descriptors extracted from training images to find a certain number of
centroids. These centroids are treated as code words or visual words, and thus form a codebook.
Each image from either testing dataset or training dataset can be further encoded to a vector with
fixed dimension using the trained codebook. The VLAD encoding contributes a key concept that
it encodes any image by a collection of code words encoding vector with a consensus dimension.
The rest of this paper is organized as follows. Section 2 reviews the previous work VLAD
algorithm, section 3 elaborates the proposed method, section 4 shows experimental results, and a
conclusion is given in section 5.
2. RELATED WORK
2.1. VLAD Algorithm
VLAD(vector of locally aggregated descriptors) is proposed by Jegou[6] in 2010. Similar as
BOF(bag-of-features)[10], VLAD aims at representing one single image by a fixed number of
feature vectors aggregated by all feature descriptors extracted from this image. Such a](https://image.slidesharecdn.com/imageretrievalusingvladwith-150109042843-conversion-gate01/85/Image-Retrieval-Using-VLAD-with-Multiple-Features-1-320.jpg)
![26 Computer Science & Information Technology (CS & IT)
representation is called a VLAD encoding for an image. Initially, VLAD takes all feature
descriptors from all training images as inputs to cluster them and find a fixed number of centroids
by K-means[7] as shown in Figure 1 and Figure 2. The collection of these centroids is referred to
as the codebook. To encode an image using the codebook, the details of processing are elaborated
as follows. Let N denote the total number of centroids, and c୧(i=1…N) denote the centroids.
Figure 1. Features extraction from images.
Figure 2. Clustering all features by K-means.
For encoding an image, we first extract all feature vectors from this image and denote them by
x୲(t=1…T) as shown in Figure 3. Then, each feature finds the centroid closest to it, NNሺx୲ሻ,
defined in eq. (1) as shown in Figure 4.
NNሺx୲ሻ = arg min ||x୲ − c୧|| (1)
Let v୧ denote the normalized vector sum of all difference between each feature vector x୲ and the
centroid NNሺx୲ሻ which it belongs to, as defined in eq. (2) and eq. (3). Then v୧(i=1…N) can be
seen as an aggregation of all feature vectors contained in the input image based on the codebook
as illustrated in Figure 5 and Figure 6., and such an aggregation is called a VLAD encoding for
this image.
v୧ = ∑ ሺx୲ − c୧ሻ୶౪:ሺ୶౪ሻୀୡ
(2)
v୧: = v୧/||v୧||ଶ (3)](https://image.slidesharecdn.com/imageretrievalusingvladwith-150109042843-conversion-gate01/85/Image-Retrieval-Using-VLAD-with-Multiple-Features-2-320.jpg)
![28 Computer Science & Information Technology (CS & IT)
3. PROPOSED METHOD
3.1. Multiple Feature Extraction
VLAD adopts SIFT as the feature descriptors for encoding. The state-of-the-art SIFT proposed by
Lowe[1] is a well-known feature descriptor which is widely applied in computer vision, object
recognition and machine intelligence due to its feature distinctness, robustness, and scale and
rotation invariant. For achieving higher accuracy for large scale image retrieval, this paper
proposes the scheme of integrating multiple types of feature into VLAD algorithm to enhance the
distinctness between image objects. Thus, besides SIFT, we also adopt the well-known feature
descriptors, including SURF, DAISY, and HOG. Detail of each feature descriptor is elaborated in
the following.
SURF is proposed by Bay[4] which replaces the DoG in SIFT with Haar wavelets transform to
generate the pyramid of scale space and approximates the determinant of Hessian blob detector
by an integer evaluation for efficiency. It is feasible to be used on large scale image matching for
the real-time concern. SURF descriptors can be extracted under various dimensions, for instance,
SURF-64 or SURF-128 are the SURF descriptors with dimension 64 or 128, respectively.
DAISY is a local dense feature descriptor scheme proposed by Tola[2]. Similar with SIFT, it
generates block-based orientation histogram but uses Gaussian convolution to aggregate these
blocks for efficiency. It has been shown that DAISY is a feasible and efficient feature descriptor
to be applied in wide base-line stereo matching. In this work, we uses dimension 200 to extract
DAISY descriptors.
The HOG descriptor is proposed by Dalal and Triggs[8], is a well-known feature descriptor and
widely applied in human or pedestrian detections due to its robustness of geometric shape and
luminance. It consists of three hierarchical structures, naming cell, block, and window,
respectively. The feature extraction starts by processing the image to a greyscale form, and
divides the image into several cells. Each cell is partitioned into nine bins according to the
orientation of gradients, and every four neighboring cells form a block. Then use a block to scan
the window for each step, the length of one cell at a time. Eventually, a feature descriptor is
generated by integrating feature vectors in all blocks. In this paper, we define that a cell consists
of 8*8 pixels, and four cells form a block with 16*16 pixels. The block then scans a window
consisting of 64*128 pixels for each step with length 8 pixels. Thus, every cell has nine
orientation features, and each block contains 36 features. After window scanning is completed,
we obtain (64-8)/8=7 scanning blocks in horizontal direction，and (128-8)/8=15 scanning blocks
in vertical direction. Therefore, the dimension of such a HOG descriptor is 64*7*15=3780.
3.2. Codebook Training
As aforementioned, VLAD uses k-means to cluster all SIFT features extracted from all training
images to find a certain number of centroids, the codebook. Similarly, we can extract other types
of feature descriptors, such as SURF-64, SURF-128, DAISY, and HOG from all training images
to compute the corresponding codebooks. Besides, the dimension of codebook, i.e., the number of
centroids for K-means clustering is also a crucial effect upon the accuracy of recognition. In this
paper, we train codebooks with 64 clusters, 128 clusters, and 256 clusters, respectively.
3.3. Combined VLAD Encoding
After finishing the codebooks generation, the encoding process for an image can be elaborated as
follows. An image can be computed for its VLAD encoding vector with a specific codebook. The
proposed encoding method is to combine several VLAD encoding vectors that are encoded by](https://image.slidesharecdn.com/imageretrievalusingvladwith-150109042843-conversion-gate01/85/Image-Retrieval-Using-VLAD-with-Multiple-Features-4-320.jpg)
![Computer Science & Information Technology (CS & IT) 29
different codebooks into a normalized encoding vector. For instance, if four types of features,
SIFT, SURF-64, DAISY and HOG are adopted and the sizes of codebook are all set to 64, an
image is firstly encoded by SIFT codebook, followed by a normalization process, to produce a
VLAD encoding vector of SIFT with dimension 128*64=8192. Applying similar processes, we
can compute the normalized VLAD encoding vector of SURF-64, DAISY, and HOG for this
image, respectively. Dimensions of SURF-64, DAISY, and HOG for a single VLAD encoding
vector are 64*64=4096, 200*64=12800, and 3780*64=241920, respectively. Thus, an image can
be represented by a combined normalized VLAD encoding vector with dimension
8192+4096+12800+241920=267008.
4. EXPERIMENTAL RESULTS
The training and testing images adopted in this research are from INRIA Holiday dataset[5]。The
Holiday dataset contains 500 classes of images, and a total of 1491 images. We pick one image
from each class for testing data, and the rest are for training data. Firstly, each training image is
pre-processed by VLAD encoding with a specific combination of codebooks. The 500 testing
image are then encoded by applying the same combination of codebook. To match the testing
image with the training images, the KNN[3] algorithm is adopted to list the first 1000 ranks of
encodings of training images for each encoding of testing image, and thus the mAP(mean average
precision) value can be computed to reflect the precision of retrieval.
Table 1 shows the performance of encodings which combines different numbers of feature
descriptors from one to four types. It is obviously that mAPs are noticeably promoted as the
number of feature types and the number of centroids increases. In all combinations of any two
types of feature descriptors as shown in table 2, the combination of SIFT and DAISY performs
best. Our explanation is that SIFT and SURF are local descriptors, while DAISY and HOG
preserves a certain portion of region description. Thus, the complementary between SIFT and
DAISY can achieve more complete and detail representation of an image, and thus make each
encoding more distinguishable from others. The combination of all four types of feature
descriptors as shown in Table 3, SIFT, SURF(64), DAISY, and HOG, with 256 centroids yields
the best result and significantly promotes the mAP to 0.72.
Table 1
Feature Descriptors Dimension
64-
centroids
mAP
Dimension
128-
centroids
mAP
Dimension
256-
centroids
mAP
SIFT 8192 0.52
8
16384 0.57
4
32768 0.60
4
SIFT+SURF(64) 12288 0.57
2
24576 0.59
7
49152 0.61
9
SIFT+SURF(64)+HOG 254208 0.63
7
508416 0.66
1
1016832 0.67
1
SIFT+SURF(64)+DAISY 25088 0.66
8
50176 0.67
9
100352 0.70
1
SIFT+SURF(64)+DAISY+HO
G
267008 0.68
7
534016 0.70
7
1068032 0.72
0](https://image.slidesharecdn.com/imageretrievalusingvladwith-150109042843-conversion-gate01/85/Image-Retrieval-Using-VLAD-with-Multiple-Features-5-320.jpg)
![Computer Science & Information Technology (CS & IT) 31
process with KNN if compared to the traditional two-stage point-to-point and geometrical
matching processes for every pair of images. The benefits of distinctiveness and efficient
matching make it practical and feasible to apply on large scale image retrieval. Although the
dimension of encoding might increase as multiple features are applied, the computing cost caused
by large dimension can be alleviated by the novel technologies of parallel processing or
distributive computing.
ACKNOWLEDGEMENT
This project is fully supported by Ministry of Science and Technology in Taiwan under grant
number 103-2815-C-143-008-E.
REFERENCES
[1] D. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91–220, 2004.
[2] Engin Tola, Vincent Lepetit, Pascal Fua. IEEE Transactions on Pattern Analysis and Machine
Intelligence. Vol. 32, Nr. 5, pp. 815 - 830, May 2010.
[3] Altman, N. S. An introduction to kernel and nearest-neighbor nonparametric regression. The
American Statistician. 1992.
[4] Herbert Bay, Tinne Tuytelaars, Luc Van Gool. Speeded Up Robust Features. ECCV, 2006.
[5] Herve Jegou, Matthijs Douze and Cordelia Schmid. Hamming Embedding and Weak geometry
consistency for large scale image search. ECCV, 2008.
[6] H.Jégou, M. Douze, C. Schmid, and P. Pérez. Aggregating local descriptors into a compact image
representation. CVPR, 2010.
[7] J.B. MacQueen. Some Methods for classification and Analysis of Multivariate Observations.
Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability, Berkeley,
University of California Press, 1967.
[8] N.Dalal and B. Triggs. Histograms of Oriented Gradients for Human Detection. CVPR, 2005.
[9] R.Arandjelovic and A. Zisserman. All about vlad. CVPR, 2013.
[10] S.Lazebnik, C. Schmid, and J. Ponce. Beyond Bags of Features: Spatial Pyramid Matching for
Recognizing Natural Scene Categories. CVPR, 2006.](https://image.slidesharecdn.com/imageretrievalusingvladwith-150109042843-conversion-gate01/85/Image-Retrieval-Using-VLAD-with-Multiple-Features-7-320.jpg)
Image Retrieval Using VLAD with Multiple Features
- 1. David C. Wyld et al. (Eds) : ITCS, CST, JSE, SIP, ARIA, DMS - 2015 pp. 25–31, 2015. © CS & IT-CSCP 2015 DOI : 10.5121/csit.2015.50103 IMAGE RETRIEVAL USING VLAD WITH MULTIPLE FEATURES Pin-Syuan Huang, Jing-Yi Tsai, Yu-Fang Wang, and Chun-Yi Tsai Department of Computer Science and Information Engineering, National Taitung University, Taiwan, R.O.C. {u10011104,u10011127,u10011139}@ms100.nttu.edu.tw; [email protected] ABSTRACT The objective of this paper is to propose a combinatorial encoding method based on VLAD to facilitate the promotion of accuracy for large scale image retrieval. Unlike using a single feature in VLAD, the proposed method applies multiple heterogeneous types of features, such as SIFT, SURF, DAISY, and HOG, to form an integrated encoding vector for an image representation. The experimental results show that combining complementary types of features and increasing codebook size yield high precision for retrieval. KEYWORDS VLAD, SIFT, SURF, DAISY, HOG 1. INTRODUCTION Image retrieval is one of the classical problems in computer vision and machine intelligence. The challenge of image retrieval is mainly aimed at trade-off between computing costs and the precision of retrieval due to the large scale data, including the large size of a single image and the large number of all images. In this paper, we propose an combinatorial encoding algorithm based on VLAD[6,9] with multiple features to achieve the goal of large scale image retrieval. VLAD cluster all features descriptors extracted from training images to find a certain number of centroids. These centroids are treated as code words or visual words, and thus form a codebook. Each image from either testing dataset or training dataset can be further encoded to a vector with fixed dimension using the trained codebook. The VLAD encoding contributes a key concept that it encodes any image by a collection of code words encoding vector with a consensus dimension. The rest of this paper is organized as follows. Section 2 reviews the previous work VLAD algorithm, section 3 elaborates the proposed method, section 4 shows experimental results, and a conclusion is given in section 5. 2. RELATED WORK 2.1. VLAD Algorithm VLAD(vector of locally aggregated descriptors) is proposed by Jegou[6] in 2010. Similar as BOF(bag-of-features)[10], VLAD aims at representing one single image by a fixed number of feature vectors aggregated by all feature descriptors extracted from this image. Such a
- 2. 26 Computer Science & Information Technology (CS & IT) representation is called a VLAD encoding for an image. Initially, VLAD takes all feature descriptors from all training images as inputs to cluster them and find a fixed number of centroids by K-means[7] as shown in Figure 1 and Figure 2. The collection of these centroids is referred to as the codebook. To encode an image using the codebook, the details of processing are elaborated as follows. Let N denote the total number of centroids, and c୧(i=1…N) denote the centroids. Figure 1. Features extraction from images. Figure 2. Clustering all features by K-means. For encoding an image, we first extract all feature vectors from this image and denote them by x୲(t=1…T) as shown in Figure 3. Then, each feature finds the centroid closest to it, NNሺx୲ሻ, defined in eq. (1) as shown in Figure 4. NNሺx୲ሻ = arg min ||x୲ − c୧|| (1) Let v୧ denote the normalized vector sum of all difference between each feature vector x୲ and the centroid NNሺx୲ሻ which it belongs to, as defined in eq. (2) and eq. (3). Then v୧(i=1…N) can be seen as an aggregation of all feature vectors contained in the input image based on the codebook as illustrated in Figure 5 and Figure 6., and such an aggregation is called a VLAD encoding for this image. v୧ = ∑ ሺx୲ − c୧ሻ୶౪:ሺ୶౪ሻୀୡ (2) v୧: = v୧/||v୧||ଶ (3)
- 3. Computer Science & Information Technology (CS & IT) 27 Figure 3. Features extraction for an image encoding. Figure 4. Each feature finds the centroid closest to it Figure 5. Distributing all features to centroids Figure 6. Vectors aggregation and normalization
- 4. 28 Computer Science & Information Technology (CS & IT) 3. PROPOSED METHOD 3.1. Multiple Feature Extraction VLAD adopts SIFT as the feature descriptors for encoding. The state-of-the-art SIFT proposed by Lowe[1] is a well-known feature descriptor which is widely applied in computer vision, object recognition and machine intelligence due to its feature distinctness, robustness, and scale and rotation invariant. For achieving higher accuracy for large scale image retrieval, this paper proposes the scheme of integrating multiple types of feature into VLAD algorithm to enhance the distinctness between image objects. Thus, besides SIFT, we also adopt the well-known feature descriptors, including SURF, DAISY, and HOG. Detail of each feature descriptor is elaborated in the following. SURF is proposed by Bay[4] which replaces the DoG in SIFT with Haar wavelets transform to generate the pyramid of scale space and approximates the determinant of Hessian blob detector by an integer evaluation for efficiency. It is feasible to be used on large scale image matching for the real-time concern. SURF descriptors can be extracted under various dimensions, for instance, SURF-64 or SURF-128 are the SURF descriptors with dimension 64 or 128, respectively. DAISY is a local dense feature descriptor scheme proposed by Tola[2]. Similar with SIFT, it generates block-based orientation histogram but uses Gaussian convolution to aggregate these blocks for efficiency. It has been shown that DAISY is a feasible and efficient feature descriptor to be applied in wide base-line stereo matching. In this work, we uses dimension 200 to extract DAISY descriptors. The HOG descriptor is proposed by Dalal and Triggs[8], is a well-known feature descriptor and widely applied in human or pedestrian detections due to its robustness of geometric shape and luminance. It consists of three hierarchical structures, naming cell, block, and window, respectively. The feature extraction starts by processing the image to a greyscale form, and divides the image into several cells. Each cell is partitioned into nine bins according to the orientation of gradients, and every four neighboring cells form a block. Then use a block to scan the window for each step, the length of one cell at a time. Eventually, a feature descriptor is generated by integrating feature vectors in all blocks. In this paper, we define that a cell consists of 8*8 pixels, and four cells form a block with 16*16 pixels. The block then scans a window consisting of 64*128 pixels for each step with length 8 pixels. Thus, every cell has nine orientation features, and each block contains 36 features. After window scanning is completed, we obtain (64-8)/8=7 scanning blocks in horizontal direction，and (128-8)/8=15 scanning blocks in vertical direction. Therefore, the dimension of such a HOG descriptor is 64*7*15=3780. 3.2. Codebook Training As aforementioned, VLAD uses k-means to cluster all SIFT features extracted from all training images to find a certain number of centroids, the codebook. Similarly, we can extract other types of feature descriptors, such as SURF-64, SURF-128, DAISY, and HOG from all training images to compute the corresponding codebooks. Besides, the dimension of codebook, i.e., the number of centroids for K-means clustering is also a crucial effect upon the accuracy of recognition. In this paper, we train codebooks with 64 clusters, 128 clusters, and 256 clusters, respectively. 3.3. Combined VLAD Encoding After finishing the codebooks generation, the encoding process for an image can be elaborated as follows. An image can be computed for its VLAD encoding vector with a specific codebook. The proposed encoding method is to combine several VLAD encoding vectors that are encoded by
- 5. Computer Science & Information Technology (CS & IT) 29 different codebooks into a normalized encoding vector. For instance, if four types of features, SIFT, SURF-64, DAISY and HOG are adopted and the sizes of codebook are all set to 64, an image is firstly encoded by SIFT codebook, followed by a normalization process, to produce a VLAD encoding vector of SIFT with dimension 128*64=8192. Applying similar processes, we can compute the normalized VLAD encoding vector of SURF-64, DAISY, and HOG for this image, respectively. Dimensions of SURF-64, DAISY, and HOG for a single VLAD encoding vector are 64*64=4096, 200*64=12800, and 3780*64=241920, respectively. Thus, an image can be represented by a combined normalized VLAD encoding vector with dimension 8192+4096+12800+241920=267008. 4. EXPERIMENTAL RESULTS The training and testing images adopted in this research are from INRIA Holiday dataset[5]。The Holiday dataset contains 500 classes of images, and a total of 1491 images. We pick one image from each class for testing data, and the rest are for training data. Firstly, each training image is pre-processed by VLAD encoding with a specific combination of codebooks. The 500 testing image are then encoded by applying the same combination of codebook. To match the testing image with the training images, the KNN[3] algorithm is adopted to list the first 1000 ranks of encodings of training images for each encoding of testing image, and thus the mAP(mean average precision) value can be computed to reflect the precision of retrieval. Table 1 shows the performance of encodings which combines different numbers of feature descriptors from one to four types. It is obviously that mAPs are noticeably promoted as the number of feature types and the number of centroids increases. In all combinations of any two types of feature descriptors as shown in table 2, the combination of SIFT and DAISY performs best. Our explanation is that SIFT and SURF are local descriptors, while DAISY and HOG preserves a certain portion of region description. Thus, the complementary between SIFT and DAISY can achieve more complete and detail representation of an image, and thus make each encoding more distinguishable from others. The combination of all four types of feature descriptors as shown in Table 3, SIFT, SURF(64), DAISY, and HOG, with 256 centroids yields the best result and significantly promotes the mAP to 0.72. Table 1 Feature Descriptors Dimension 64- centroids mAP Dimension 128- centroids mAP Dimension 256- centroids mAP SIFT 8192 0.52 8 16384 0.57 4 32768 0.60 4 SIFT+SURF(64) 12288 0.57 2 24576 0.59 7 49152 0.61 9 SIFT+SURF(64)+HOG 254208 0.63 7 508416 0.66 1 1016832 0.67 1 SIFT+SURF(64)+DAISY 25088 0.66 8 50176 0.67 9 100352 0.70 1 SIFT+SURF(64)+DAISY+HO G 267008 0.68 7 534016 0.70 7 1068032 0.72 0
- 6. 30 Computer Science & Information Technology (CS & IT) Table 2 Feature Descriptors Dimension 64- centroids mAP Dimension 128- centroids mAP Dimension 256- centroids mAP DAISY+HOG 254720 0.52 7 509440 0.52 6 1018880 0.53 8 SURF(128)+HOG 250112 0.56 6 500224 0.59 8 1000448 0.58 5 SURF(64)+HOG 246016 0.56 7 492032 0.58 5 984064 0.59 8 SIFT+SURF(128) 16384 0.57 1 32768 0.59 0 65536 0.60 7 SIFT+SURF(64) 12288 0.57 2 24576 0.59 7 49152 0.61 9 SURF(128)+DAISY 20992 0.59 4 41984 0.60 8 83968 0.60 9 SURF(64)+DAISY 16896 0.60 1 33792 0.60 7 67584 0.62 6 SIFT+HOG 250112 0.60 4 500224 0.64 6 1000448 0.65 8 SIFT+DAISY 20992 0.62 4 41984 0.64 1 83968 0.65 5 Table 3 Feature Descriptors Dimension 64- centroids mAP Dimension 128- centroids mAP Dimension 256- centroids mA P SIFT+SURF(128)+HOG 258304 0.63 0 516608 0.66 6 1033216 0.66 2 SURF(64)+DAISY+HOG 258816 0.63 0 517632 0.64 2 1035264 0.65 4 SURF(128)+DAISY+HOG 262912 0.63 5 525824 0.64 7 1051648 0.64 5 SIFT+SURF(64)+HOG 254208 0.63 7 508416 0.66 1 1016832 0.67 1 SIFT+DAISY+HOG 262912 0.64 1 525824 0.66 7 1051648 0.67 5 SIFT+SURF(128)+DAISY 29184 0.66 1 58368 0.68 0 116736 0.68 6 SIFT+SURF(64)+DAISY 25088 0.66 8 50176 0.67 9 100352 0.70 1 SIFT+SURF(64)+DAISY+HO G 267008 0.68 7 534016 0.70 7 1068032 0.72 0 SIFT+SURF(128)+DAISY+H OG 271104 0.69 2 542208 0.71 1 1084416 0.70 7 5. CONCLUSIONS In this paper, we show that a combinatorial encoding method with multiple features indeed enhance the distinctiveness among large number of image representations, and thus significantly promote the retrieval precision. On the other hand, encoding by using identical size of codebook equalizes the dimension of representation for each image, and efficiently facilitates the matching
- 7. Computer Science & Information Technology (CS & IT) 31 process with KNN if compared to the traditional two-stage point-to-point and geometrical matching processes for every pair of images. The benefits of distinctiveness and efficient matching make it practical and feasible to apply on large scale image retrieval. Although the dimension of encoding might increase as multiple features are applied, the computing cost caused by large dimension can be alleviated by the novel technologies of parallel processing or distributive computing. ACKNOWLEDGEMENT This project is fully supported by Ministry of Science and Technology in Taiwan under grant number 103-2815-C-143-008-E. REFERENCES [1] D. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 60(2):91–220, 2004. [2] Engin Tola, Vincent Lepetit, Pascal Fua. IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 32, Nr. 5, pp. 815 - 830, May 2010. [3] Altman, N. S. An introduction to kernel and nearest-neighbor nonparametric regression. The American Statistician. 1992. [4] Herbert Bay, Tinne Tuytelaars, Luc Van Gool. Speeded Up Robust Features. ECCV, 2006. [5] Herve Jegou, Matthijs Douze and Cordelia Schmid. Hamming Embedding and Weak geometry consistency for large scale image search. ECCV, 2008. [6] H.Jégou, M. Douze, C. Schmid, and P. Pérez. Aggregating local descriptors into a compact image representation. CVPR, 2010. [7] J.B. MacQueen. Some Methods for classification and Analysis of Multivariate Observations. Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability, Berkeley, University of California Press, 1967. [8] N.Dalal and B. Triggs. Histograms of Oriented Gradients for Human Detection. CVPR, 2005. [9] R.Arandjelovic and A. Zisserman. All about vlad. CVPR, 2013. [10] S.Lazebnik, C. Schmid, and J. Ponce. Beyond Bags of Features: Spatial Pyramid Matching for Recognizing Natural Scene Categories. CVPR, 2006.