Keiji Gyohten

An evaluation on robustness and brittleness of HOG features of human detection

Detecting humans in an image sequence is one of the most difficult problems in object recognition. It is necessary to define a robust descriptor which can extract human features from images, to improve the detecting performance. Histograms of Oriented Gradients(HOG) descriptor significantly outperforms compared with the others on human detection. The descriptor is known as a robustness descriptor for illumination changes and geometrical changes in local regions. To obtain the high detection performance using the descriptor effectively, it is necessary to know the robustness and brittleness of the descriptor. In this paper, we experiment the descriptor to verify its robustness to illumination changes and to scrutinize its brittleness. For the experiments, we use LogitBoost which can create a human-detector by learning human features.

2D CAD Data Mining Based on Spatial Relation

In this research, we propose CAD data mining technique to obtain semantic elements without prior knowledge about plans being designed. Our method consists of two steps. The first step is to extract frequent spatial relations between figure elements in CAD data as clues to the semantic elements. These relations are modeled as topology graph and are analyzed by a graph mining method. In the second step, valid semantic elements are specified by eliminating geometrically unnecessary figure elements through inferring every affine transformation between sets of figure elements having the same frequent spatial structure. In the experiments, the proposed method could extract semantic elements like electrical symbols from floor plan data without prior knowledge about the symbols.

A self-diagnosis under 2D projectivity for local descriptor base template matching

2D projectivity is an invertible mapping to present the perspective imaging of a world plane by projective translation, called homography. Good image feature have to be robust under 2D projectivity caused by any camera movements. In the standard performance evaluation of template matching, many real captured images of many scenes are ordinarily used. However it is not enough to evaluate the robustness under 2D projectivity in detail because the variations of real camera pose and position in the 3D world are limited and the capturing cost is expensive. During the early stage of the template matching development, an easy performance evaluation method is required to examine the behavior. We propose a self-diagnosis method to measure the robustness of local descriptor base template matching between a template image and reference images which are created by projective translation of the template image. We focus on the template matching consisting of a feature point extraction and a local descriptor matching. The proposed method evaluates the spatial accuracy of the feature points and the estimated template positions in the reference images with local descriptor matchings. Four metrics, feature point precision (PP), feature point recall (PR), local descriptor matching precision (MP) and local descriptor matching recall (MR) are introduced to evaluate the performance. The experiment results will be appeared in the final manuscript to show the effectiveness of our method.

2D Figure Pattern Mining

1.1 Background With the recent enhancement of desktop design environments, it has become easy for personal users to design graphical documents such as posters, flyers, slides, drawings, etc. These kinds of documents are usually produced by the applications like drawing softwares, which have the advantage that they can store and retrieve the drawing data electronically. By reusing parts of the stored drawing data, the users can design the graphical documents much more easily. However, generally, the stored data of many users is not shared, although this can be achieved by putting a drawing database. One reason is that it is difficult to retrieve desired figures from large amounts of drawing data in the database. Unlike in text search, the figure search will require enormous amounts of computation time because matching of the geometric primitives in the drawing data will cause their combinatorial explosion in 2D space. To address this problem, many approaches have been proposed recently. When users search the drawing database, they should conjure up the desired figures and design their 2D sketches as the keys. In case of retrieving general figures, such as electrical symbols and map symbols, there would be little difference between sketches of them drawn by different users, but it is impractical to make them visualize various objects and things and use the sketches as the keys. For example, in case of retrieving human figures, since the sketches of humans differ according to the users, not all of figures of humans will be able to be obtained from the drawing database. To cope with this problem, we need a technique that enables the applications to automatically present users with the list of figures considered to have any meaning. Users can specify a figure of the desired object or thing simply by selecting it from the list and then retrieve the desired figures from the database using it as the key.

An Impressive Data Animating for Positional Movement of Soccer Players

In this paper, we propose a system which decides camera work in a 3D virtual soccer game and generates 3DCG animation automatically, called data animating. The 3D virtual soccer game is represented by the play events including player IDs, player positions, types of play and so on. These properties are created by real soccer video analysis system, called digital scorebook. The system generates a 3DCG animation according to two user selections of soccer play events. To create impressive soccer animation from real soccer game, we focus on virtual cinematography as the art of rhetoric of animation scene. The system can automate a composition of animation story and decide camera work based on the virtual cinematography. Our animation story has the three part, introduction, development and conclusion, selected from the organization of Chinese poetry. The appropriate camera work is decided in each part. Moreover, the picture composition for each part in a screen is calculated based on golden ratio. In the experimental result, the system can generate 3DCG animation automatically from the play events selected by user

Community partition based chat system easy to join for new members

In a chat system, an increase of the participants leads to confusion of topics and makes it difficult to continue an orderly conversation. Moreover, senior participants tend to be familiar with each other and form a so-called community. This behavior of the senior users consequently makes newcomers hard to participate in the existing rooms. To cope with these problems, this paper proposes a community partition based chat system which divides overcrowded rooms to keep the proper number of participants. The system frequently estimates relationship between participants and separates the community members from the chat room. The separated members thus can continue to communicate with each other in another room newly generated, while the remained participants unacquainted with each other can make a beginning of the conversation. We had made an evaluation experiment on the proposed chat system and could obtain results which agree with the aim of our research

A color feature learning and robust interpretation of moving object using HMM

Spot observation by computer vision is the one of fundamental key technology. In this paper, we propose a moving object color learning and robust recognition with Hidden Markov Model(HMM) from various scenes under different light conditions. Feature box which is a small area in a image is defined to observe a spot. The time series data of such as averages of R, G, B intensities in feature boxes are the input signals of our system. The HMMs learn correspondences of input signals with object color of moving object and background. Baum-Welch and Vi-terbi algorithms are used to learning and interpret the spot scene transition. In moving object color interpretation, the system selects a best HMM model for input signals using maximum likelihood method based on a given object color appearance grammar. In the experiment, we examine the number of feature boxes and its shapes under some light conditions. The feature boxes adjoining in vertical column whose height is almost same as objects results best score in the experiment. It shows the effectiveness of our method.

Optimization-based image analysis dealing with symbolic constraints using hierarchical multi-agent system

The paper describes a method for understanding an image where desired objects have part-of relationships between them. This method is based on a hierarchical multi-agent system, where each agent takes charge of a desired object and tries to extract it using knowledge on its features. Since users can define this knowledge freely without any modification of the algorithm, this method is applicable to various problems of image analysis by changing the knowledge. Moreover, the agents in this system use symbolic constraints and evaluation measurements on the desired objects. They are defined in the knowledge each agent has and used to obtain the desired results where obtained objects are evaluated highly in terms of the evaluation measurements and satisfy their plausible relationships defined symbolically. To verify our method experimentally, we applied it to problems of line drawing recognition and character segmentation.