Mitsuru Ambai

CARD: Compact And Real-time Descriptors

We propose Compact And Real-time Descriptors (CARD) which can be computed very rapidly and be expressed by short binary codes. An efficient algorithm based on lookup tables is presented for extracting histograms of oriented gradients, which results in approximately 16 times faster computation time per descriptor than that of SIFT. Our lookup-table-based approach can handle arbitrary layouts of bins, such as the grid binning of SIFT and the log-polar binning of GLOH, thus yielding sufficient discrimination power. In addition, we introduce learning-based sparse hashing to convert the extracted descriptors to short binary codes. This conversion is achieved very rapidly by multiplying a very sparse integer weight matrix by the descriptors and aggregating signs of their multiplications. The weight matrix is optimized in a training phase so as to make Hamming distances between encoded training pairs reflect visual dissimilarities between them. Experimental results demonstrate that CARD outperforms previous methods in terms of both computation time and memory usage.

Keypoint detection by cascaded fast

When the FAST method for detecting corner features at high speed is applied to images that include complex textures (regions that include foliage, shrubbery, etc.), many corners that are not needed for object recognition are detected because FAST defines corner features on the basis of a 16-pixel bounding circle. To overcome that problem, we propose the Cascaded FAST that defines corners on the basis of similarity in terms of intensity, continuity and orientation in a broader range of areas (20, 16, and 12 pixel bounding circles). Also, cascading three decision trees trained by the FAST approach enables high-speed corner detection in which non-corners are eliminated early in the process. Furthermore, Cascaded FAST determines scale by using an image pyramid and determines orientation at high speed by using a framework for referencing surrounding pixels.

Multiclass VisualRank: image ranking method in clustered subsets based on visual features

This paper proposes Multiclass VisualRank, a method that expands the idea of VisualRank into more than one category of images. Multiclass VisualRank divides images retrieved from search engines into several categories based on distinctive patterns of visual features, and gives ranking within the category. Experimental results show that our method can extract several different image categories relevant to given keyword and gives good ranking scores to retrieved images.

SPADE: Scalar Product Accelerator by Integer Decomposition for Object Detection

We propose a method for accelerating computation of an object detector based on a linear classifier when objects are expressed by binary feature vectors. Our key idea is to decompose a real-valued weight vector of the linear classifier into a weighted sum of a few ternary basis vectors so as to preserve the original classification scores. Our data-dependent decomposition algorithm can approximate the original classification scores by a small number of the ternary basis vectors with an allowable error. Instead of using the original real-valued weight vector, the approximated classification score can be obtained by evaluating the few inner products between the binary feature vector and the ternary basis vectors, which can be computed using extremely fast logical operations. We also show that each evaluation of the inner products can be cascaded for incorporating early termination. Our experiments revealed that the linear filtering used in a HOG-based object detector becomes 36.9× faster than the original implementation with 1.5% loss of accuracy for 0.1 false positives per image in pedestrian detection task.

Sparse Isotropic Hashing

This paper address the problem of binary coding of real vectors for efficient similarity computations. It has been argued that orthogonal transformation of center-subtracted vectors followed by sign function produces binary codes which well preserve similarities in the original space, especially when orthogonally transformed vectors have covariance matrix with equal diagonal elements. We propose a simple hashing algorithm that can orthogonally transform an arbitrary covariance matrix to the one with equal diagonal elements. We further expand this method to make the projection matrix sparse, which yield faster coding. It is demonstrated that proposed methods have comparable level of similarity preservation to the existing methods.

Distance Computation Between Binary Code and Real Vector for Efficient Keypoint Matching

Image recognition in client server system has a problem of data traffic. However, reducing data traffi cg ives rise to worsening of performance. Therefore, we represent binary codes as high dimensional local features in client side, and represent real vectors in server side. As a result, we can suppress the worsening of the performance, but it problems of an increase in the computational cost of the distance computation and a different scale of norm between feature vectors. Therefore, to solve the first problem, we optimize the scale factor so as to absorb the scale difference of Euclidean norm. For second problem, we compute efficiently the Euclidean distance by decomposing the real vector into weight factors and binary basis vectors. As a result, the proposed method achieves the keypoint matching with high-speed and high-precision even if the data traffic was reduced.

Fast and Accurate Object Detection Based on Binary Co-occurrence Features

In this paper, we propose a fast and accurate object detection algorithm based on binary co-occurrence features. In our method, co-occurrences of all the possible pairs of binary elements in a block of binarized HOG are enumerated by logical operations, i.g. circular shift and XOR. This resulted in extremely fast co-occurrence extraction. Our experiments revealed that our method can process a VGA-size image at 64.6 fps, that is two times faster than the camera frame rate (30 fps), on only a single core of CPU (Intel Core i7-3820 3.60 GHz), while at the same time achieving a higher classification accuracy than original (real-valued) HOG in the case of a pedestrian detection task.

Asymmetric Feature Representation for Object Recognition in Client Server System

This paper proposes asymmetric feature representation and efficient fitting feature spaces for object recognition in client server system. We focus on the fact that the server-side has more sufficient memory and computation power compared to the client-side. Although local descriptors must be compressed on the client-side due to the narrow bandwidth of the Internet, feature vector compression on the server-side is not always necessary. Therefore, we propose asymmetric feature representation for descriptor matching. Our method is characterized by the following three factors. The first is asymmetric feature representation between client- and server-side. Although the binary hashing function causes quantization errors due to the computation of the sgn function \((\cdot )\), which binarizes a real value into \(\{1,-1\}\), such errors only occur on the client-side. As a result, performance degradation is suppressed while the volume of data traffic is reduced. The second is scale optimization to fit two different feature spaces. The third is fast implementation of distance computation based on real-vector decomposition. We can compute efficiently the squared Euclidean distance between the binary code and the real vector. Experimental results revealed that the proposed method helps reduce data traffic while maintaining the object retrieval performance of a client server system.