Presented by Xinyu Chang Cross-Indexing of Binary Scale Invariant Feature Transform Codes for Large-Scale Image Search Presented by Xinyu Chang

Introduction Image matching is a fundamental aspect of many problems in computer vision, including object or scene recognition, solving for 3D structure from multiple images, stereo correspondence, and motion tracking. In recent years, there has been growing interest in mapping visual features into compact binary codes for applications on large-scale image collections. Encoding high-dimensional data as compact binary codes reduces the memory cost for storage.

Introduction Goal Extracting distinctive invariant features Correctly matched against a large database of features from many images Invariance to image scale and rotation Robustness to Affine distortion Change in 3D viewpoint Addition of noise Change in illumination

Introduction

Content Interest Point Detection Flexible Binarization Cross Indexing Scale-space extrema detection Keypoint localization Orientation assignment Keypoint descriptor Flexible Binarization Cross Indexing Result

Interest Point Detection

Interest Point Detection

Interest Point Detection

Interest Point Detection

Initial Outlier Rejection Dog is most stable across scale

Interest Point Detection

Rotation invariance To achieve rotation invariance Compute central derivatives, gradient magnitude and direction of L (smooth image) at the scale of key point (x,y)

Rotation invariance

Rotation invariance

Rotation invariance

Key point descriptor

FLEXIBLE SIFT BINARIZATION Given an image, the detected interest points are denoted by { fi }n−1 i=0 , in which N represents the total number of the detected interest points. Each feature fi includes a L2-normalized descriptor di ∈ RD, for SIFT descriptor D is 128. Our target is to transform local feature descriptor di to an L-bit binary code string B = {b0, b1, . . . , bL−1}

FLEXIBLE SIFT BINARIZATION D where C represents the 3-D comparison array with size D × D × 2. And C(i, j ) means the comparison result between the magnitudes in the i -th and the j -th dimension of descriptor d. α is a scalar threshold whose impact will be studied in the experiment section.

FLEXIBLE SIFT BINARIZATION And concatenate them into a comparison string S with β = 2D(D − 1) bits in total, as shown by the second step in Fig. 2. To simplify the notations, in the following, S is denoted as S = {s0, s1, s2, . . . , sβ−1}. To obtain an L-bit binary code B = {b0, b1, . . . , bL−1}, next we encode the comparison string S into L bits.

FLEXIBLE SIFT BINARIZATION

CROSS-INDEXING STRATEGY Code Word the first 32 bits of the binary code is code word. The visual words are generated by clustering the randomly selected SIFT descriptor. Each feature is assigned to a visual word by nearest neighbor approach or approximate nearest neighbor approach.

CROSS-INDEXING STRATEGY In the BoVW model, an image is represented by a visual word histogram with tf -idf weighting strategy. The similarity between two images are measured by the L1 or L2 distance of their visual word vectors. In the binary code based retrieval system, the features’ binary codes are used to find the true matches and we use the number of matches to measure the similarity between two images, denoted by Scorei. And this strategy can be formulated by in which i represents the i -th database image. B(d) and B(q) denote the binary SIFT code of the database feature d and the query feature q, respectively. T is a pre-defined threshold value. The impact of T will be studied in our experimental part. H(・, ・) denotes the Hamming distance between two binary SIFT codes. If two images have the same score value, we favor the image with fewer features.

CROSS-INDEXING STRATEGY

CROSS-INDEXING STRATEGY

Result

Result

Thank you