Interleaved Intensity Order Based Local Descriptor for Image Matching IEEE Transactions on Image Processing, 2014 Shiv Ram Dubey, Satish Kumar Singh and Rajat Kumar Singh Indian Institute of Information Technology, Allahabad The effect of proposed scheme over descriptor dimension is depicted in Fig. 3.
Introduction Designed an Interleaved Intensity Order based Local Descriptor (IOLD) for Local Image Matching. A generalization of the Local Intensity Order Pattern (LIOP) [1]. Inherently rotation and illumination invariant. Solved the dimensionality problem of LIOP.
10
10
Length of the Pattern
Method The local neighborhood is divided into multiple interleaved neighboring sets as shown in Fig. 1. X k 1 i
X k 2 i
X k 1 i
X k i
X 2k i
X
X N k 1 i
i
Y
X 2 i
X N i
X
O
X
X N k 1 i
Y
X N k 2 i
X
(b) Neighboring set 1
O
6
10
4
10
2
10
i
...
X 2 i
X k i
X 2k i
X
Y
X N i
X
O
(c) Neighboring set 2
X
(d) Neighboring set k
Fig.1. Considering local neighborhood as a set of different interleaved local neighborhood. The original N neighbors are divided into k neighboring sets having d=N/k neighbors each.
10
0
2 4 6 8 10 Number of neighboring sample points (N)
12
Fig.3. Comparison between the pattern dimension using LIOP [1] and proposed approach.
i
X1 i
O
(a) Neighbors of Xi
X
i
10
𝑋𝑖3
𝑋𝑖
Oxford Image Matching Dataset [2] and CASIA’s Complex Illumination Change Dataset [3] are used to test the performance and robustness of the proposed descriptor.
𝑋𝑖
𝑋𝑖4 𝑋𝑖5
𝑋𝑖6
𝑋𝑖7
O
(c)
(d)
(e)
(f)
Fig.6. Descriptors performance for kd=14, 24, 15, 25 and 16 (i.e. number of interleaved set & number of neighbors in each set) when B=1 (i.e. number of multi-scale regions) and C=1 (i.e. number of sub regions) over Oxford dataset for sequence a) leuven, b) bikes, c) ubc, d) boat, e) graf, f) wall, and g) the matching time.
℮ 𝑋𝑖 = 𝐼𝑋1𝑖 , 𝐼𝑋2𝑖 , 𝐼𝑋3𝑖 , 𝐼𝑋4𝑖 , 𝐼𝑋5𝑖 , 𝐼𝑋6𝑖 , 𝐼𝑋7𝑖 , 𝐼𝑋8𝑖 = (10,8,20,158,200,30,50,34)
𝑋𝑖8
(b)
(g)
(b) 8 Local Neighbors of the Patch
𝑋𝑖1
(a)
Datasets Used
The computation of IOLD pattern for a particular pixel is depicted in Fig. 2. The weighting factor is computed by finding the number of pairs having difference of more than 5 and biased by 1 to avoid multiplication by 0. (a)2 Example Patch
The image matching results in terms of recall vs 1-precision are depicted in Fig.6 over each sequence of Oxford Dataset. The average result over both sequence of CASIA’s Dataset is demonstrated in Fig.7. It is observed that using interleaved order the performance is improved either improved or nearly equal with a great improvement in matching time.
0
Y
X1 i
X N k 2 i
X k 2 i
LIOP pattern our pattern with k=2 our pattern with k=3 our pattern with k=4
8
Image Matching Results
(c) Interleaved Orders ℮1 𝑋𝑖 = 𝐼𝑋1𝑖 , 𝐼𝑋3𝑖 , 𝐼𝑋5𝑖 , 𝐼𝑋7𝑖 = 10,20,200,50
℮2 𝑋𝑖 = 𝐼𝑋2𝑖 , 𝐼𝑋4𝑖 , 𝐼𝑋6𝑖 , 𝐼𝑋8𝑖 = 8,158,30,34
𝑂𝑟𝑑𝑒𝑟 = (1,2,4,3)
𝑂𝑟𝑑𝑒𝑟 = (1,4,2,3)
(d) Ordering Patterns 𝐼𝑛𝑑𝑒𝑥 = 2
𝐼𝑛𝑑𝑒𝑥 = 5
24
𝜉𝑃 (℮1 𝑋𝑖 = (0,1,0,0,0,0,0, … ,0)
24
𝜉𝑃 (℮2 𝑋𝑖 = (0,0,0,0,1,0,0, … ,0)
(e) Weighted Ordering Patterns 𝕎 ℮1 𝑋𝑖 = 7 𝔓1 𝑋𝑖 = 𝕎 ℮1 𝑋𝑖 × 𝜉𝑃 ℮1 𝑋𝑖 = (0,7,0,0,0,0,0, … ,0)
𝕎 ℮2 𝑋𝑖 = 6 𝔓2 𝑋𝑖 = 𝕎 ℮2 𝑋𝑖 × 𝜉𝑃 ℮2 𝑋𝑖 = (0,0,0,0,6,0,0, … ,0)
Fig.4. Oxford image matching dataset [2]: the images in first to sixth row are having the Illumination (leuven), Image Blur (bikes), JPEG Compression (ubc), Scale and Rotation (boat), Viewpoint change (graf), and Viewpoint change (wall) effects respectively.
References
(f) Final Pattern 𝔉𝔓 𝑋𝑖 = [𝔓1 𝑋𝑖 , 𝔓2 𝑋𝑖 ] = (0,7,0,0,0,0,0, … ,0,0,0,0,0,6,0,0, … ,0)
Fig.2. (a) An example patch, (b) 8 local neighbors of a pixel selected in a rotation invariant manner, (c) Interleaved orders over 2 neighboring sets, (d) Ordering patterns, (e) Ordering patterns weighted by the local dissimilarity, and (f) Final IOLD pattern.
Fig.7. Comparison of IOLD with LIOP [1], SIFT [4] and HRI-CSLTP [5] over Complex illumination change dataset in terms of (a) recallprecision and (b) matching time.
Fig.5. Complex illumination change dataset [3].
[1]Z. Wang, B. Fan, F. Wu, Local Intensity Order Pattern for feature description, ICCV (2011) 603-610. [2]http://www.robots.ox.ac.uk/~vgg/research/affine/. [3]http://vision.ia.ac.cn/Students/wzh/datasets/illumination/Illumination_Datasets.zip. [4]D.G. Lowe, Distinctive image features from scale-invariant keypoints, IJCV 60 (2) (2004) 91–110. [5]R. Gupta, H. Patil, A. Mittal, Robust order-based methods for feature description, CVPR (2010) 334 –341.
