In this report, we propose an algorithm of multiple persons tracking by cooperative multiple cameras, in a situation that it is difficult for single camera to track multiple persons. For example, when tracking two persons with two cameras for tracking, two persons can be tracked by exchanging pictures between the cameras for tracking. However, when the number of persons or cameras for tracking increases, we encounter such a problem that the exchanging images cause the increase of the amount of communications. Then, we use the center camera for understanding a status of cameras for tracking and the position of each person. Moreover, we try to track more persons than the number of cameras for tracking.
We describe a novel system which realizes intuitive physical retrieval of 3D meshes from a database of 3D models, based not only on the shape geometry but also on its functionality. By using point-sampled representation for a 3D shape both of a query and models, our system allows the user to physically and interactively express a geometrical and content based query by simply assembling a set of ActiveCubes. The user can retrieve 3D models from the database based on their shape, and can also physically integrate functional annotated blocks in the query. Following, the user intuitively defines an interactive spatial query which has both geometrical and functional aspects.
The purpose of this paper is to derive a formula for estimating impressions from feature of landscape pictures. The formulas were applied to develop a search engine of the pictures. Then the power of the engine was evaluated by subjective rating test. The new estimated formulas were drawn from the composition and difference of brightness of color to improve the accuracy of retrieval. The new formulas were compared with the formulas draw from picture composition to examine the validity of the search engine.
The Mahalanobis distance was newly introduced into the decision of pitching scenes on a video stream of baseball games. This new method increased by a factor of 5 to 10% the recall/precision rates of comparison between the reference key and the specific shots picked up from a video stream, as compared with the conventional Watabe-Komiya method. The new method is realized by Mahalanobis distance applied to the realtime calculus of the segmented motion vector magnitudes on the DSP algorithms. When the specific shots was experimentally designated realtime from the live video stream of baseball games, the actual recall rate was as high as 84 to 87%. The proposed video retrieval strategy was confirmed to be applied to detect home runs, hits, etc.
An automatic measurement system of surface shape using grating projection method has been proposed. The system is composed of an optical spatial modulator from which grating patterns are projected on the surface of the object, a CCD camera and a computer. The patterns on the surface of the object were taken into the computer by the CCD camera, and the 3D coordinate of the surface of the object was calculated. The system with high measuring speed and high accuracy was improved using the optical spatial modulator.
In additive color mixture, n (≥2 ) colors are given then the color mixture result is determined uniquely. On the other hand, in subtractive color mixture, the result is not unique which are not clarified also in today. In No.I article of the series, we first derived theorems related to the upper and lower bounds of the subtractive color mixture. Second, using these theorems, theorems related to the centroid of the possible range of a stimulus value are derived. In this article, it is shown that the centroid of resultant tristimulus values is represented approximately in the multiplication form between component tristimulus values. This multiplication corresponds to the addition in additive color mixture and has a significant meaning in the color mixture systematization.
In previous publications, we have proposed a MPEG-7 based annotation architecture, Dozen Dimensional Digital Content (DDDC), specially designed for facilitating the managing, browsing and retrieving of personal digital photograph collections. DDDC annotates multimedia data with twelve main attributes regarding its semantic representations. However, given the proposed DDDC architecture, manual annotations are still inevitable while lack of pre-attached machine generated metadata. When dealing with a large collection of personal photographs, this process could be frustrating and time consuming. To provide a more facilitated methodology, we utilize the concept of "Spatial and Temporal Based Ontology" to enable semi-automatic semantic DDDC annotation. Our "Spatial and Temporal Based Ontology" are designed based on the special pattern of personal photograph collections while we argue that time and location are two most important attributes in terms of personal photograph retrieval. This paper explains our approach.
This paper describes a process of the affine reconstruction from multiple view images, stressing on the Necker Reversal problem and a surface representation of hollow part of object. In order to obtain 3D position of the corresponding points, fundamental matrices are estimated from the corresponding points between each three images. Rotation matrix and camera direction are calculated using the fundamental matrices. The right solution of the equation is chosen according to numbers of polygon seen on images through shape reconstruction in every two solutions. The 3D points calculated from the each three images are integrated into one coordinate. The shape reconstruction from the set of 3D points is performed by removing isolated trigonal pyramid for correct surface representation of hollow part of object. In the experiments, shape of object having curved surface with hollow part are represented.
In order to evaluate halftoned images or halftoning techniques, inverse halftoning schemes of error diffused images have been studied. Most of the conventional methods are either simple filtering, which is not satisfactory, or complicated edge direction detection, which can be too complicated to design. Here we propose a method that is based on neighboring halftoned pixel pattern conditioning. It was found that the improvement of SNR of restored halftone image obtained by the proposed method, compared to the conventional 5x5 Gaussian method, is from 2 to 4dB as great.
We propose a photometric method that simultaneously estimates the surface reflectance and geometry properties from four images. In an iterative framework, the specular reflectance is estimated with a reflectance recovery algorithm which uses the data all over the object, and addresses the problem of data inadequacy. Using the result of the specular reflectance recovery, a Photometric Stereo procedure is applied to estimate the local surface parameters at local points. Our approach integrates reflectance recovery and the photometric stereo technique to yield accurate separation, reflectance and surface orientation as demonstrated by our results on synthetic and real images.