Tsuyoshi MIGITA

Hiroshima City University, Hirosima, Hiroshima, Japan

Are you Tsuyoshi MIGITA?

Claim your profile

Publications (15)0 Total impact

  • IASTED International Conference Signal and Image Processing, Honolulu; 08/2004
  • Source
    Akira Amano, Tsuyoshi Migita, Naoki Asada
    [Show abstract] [Hide abstract]
    ABSTRACT: The linearized approach to the shape from motion problem, e.g. the factorization method, is robust to find a unique solu- tion with fast computation, but the occlusionand perspective distortion are out of scope in the linear formulation. In con- trast, the nonlinear approach is free from such limitations, yet it involves two problems; one is to find the globally opti- mal solution and the other to reduce the computation time. In this paper, we present an effective nonlinear optimiza- tion method to recover 3D shape and motion. To overcome the shortcomings of the nonlinear scheme, we propose "dou- ble search (DS) procedure" and "preconditioned conjugate gradient (PCG) algorithm". The DS procedure enables us to find two major solutions that correspond to the true and false shapes and then we select the globally optimal solu- tion by evaluating the error of them. The PCG algorithm is an improved CG one whose computational performance is several times faster than the conventional Levenberg- Marquardt (LM) algorithm. We carried out experiments with simulation and real data, and the results have demonstrated that the proposed method allows us to obtain the correct shape and motion with 3-9 times faster computation than the LM algorithm. Finally we have shown the applicability of the method to a large building reconstruction from a set of partially tracked feature points.
    05/2002;
  • Source
    Masashi Baba, Naoki Asada, Ai Oda, Tsuyoshi Migita
    [Show abstract] [Hide abstract]
    ABSTRACT: Depth recovery is a central concern in computer vision, and many methods were proposed for the monocular depth esti-mation by zooming as well as focusing and irising. In the past, there are two distinct approaches in depth by zooming; one is from motion parallax along the optical axis using a pinhole camera model, and the other from defocus using a thin lens camera model. This paper presents a new camera model that accounts for both effects of defocus and lens cen-ter translation by zooming. We first discuss the optical prop-erties of zoom lenses, then present a thin lens based camera model that describes the mutual relationship between zoom, focus and iris parameters. Using this model with calibra-tion results, we have performed some experiments with real images and evaluated the accuracy of the depth information recovered from defocus and lens center translation. Experi-mental results have demonstrated the validity of our camera model and also shown its applicability to the depth estima-tion from defocus and translation by zooming.
    05/2002;
  • Source
    Tsuyoshi Migita, Akira Amano, Naoki Asada
    Proceedings of the IAPR Conference on Machine Vision Applications (IAPR MVA 2002), December 11-13, 2002, Nara-ken New Public Hall, Nara, Japan; 01/2002
  • Source
    Akira Amano, Tsuyoshi Migita, Naoki Asada
    [Show abstract] [Hide abstract]
    ABSTRACT: Fact,orization is one of the most practical lnethocl to recover 3D shape arid motion simultaneously fro111 2D images with stable and fast computation. How- ever, there still remain two crucial problems to be solved in real situations; one is to determine the true shape from a pair of visually equivalent candidates and the other is to measure the actual size of the object. This paper presents a method to solve the shape and size problems by the factorization with action; that is, 3D recovery is performed from an image sequence with a known trajectory of a sin- gle feature point given by the compiiter-controlled robot hand, arid we determine the shape and size by evaluating the co~isistency between computed shape and the given trajectory. Experi~nents results per- formed in sir nu la ti or^ study and in real world have shown the effectiveness of our niethocl.
    Proceedings of the IAPR Conference on Machine Vision Applications (IAPR MVA 2000), November 28-30, 2000, Tokyo, Japan; 01/2000
  • [Show abstract] [Hide abstract]
    ABSTRACT: 物体の全周囲を撮影した数十枚の疎な全周囲画像列から数十万点で構成される密な3次元形状モデルを生成する手法について述べる.本手法ではまず,物体を撮影した全周囲画像列から特徴点に基づく手法を用いて形状とカメラパラメータの自動推定を行う.次に,推定されたカメラパラメータと部分画像列に対してマルチベースラインステレオ法を適用し,密な部分モデルを生成する.最後に複数の部分モデルを統合することで,密な全周囲の3次元形状モデルを生成する.本手法を用いて,実物体を撮影した疎な全周囲画像列から密な3次元形状モデルが生成できることを示す. / This paper describes a method for generating three-dimensional dense model from a sparse image sequence taken around an object. First, the shape recovery and camera parameter estimation are performed based on the feature point correspondences between images. Second, several partial dense models are reconstructed by using the multiple baseline stereo method. Lastly, the entire three-dimensional dense model is generated by integrating the partial models. Experimental results have shown that the proposed method works well with real image sequences.
  • [Show abstract] [Hide abstract]
    ABSTRACT: 運動からの形状復元問題は,非線形最小2乗法による再投影誤差最小化として定式化できる.一意の解が得られる線形の定式化も知られているが,再投影誤差最小化の方が適用範囲が広く精度も高い等の利点がある.しかし,非線形問題であるたや計算量の削減(高速化)と局所解の回避(安定化)が重要な課題である.本研究では,これらの課題を扱う.また,屋外建物の全周形状復元問題を対象とし,画像数数百枚,特徴点数数百点の規模の実験を行い,提案手法の有効性を示す. / Simultaneous recovery of shape and motion from an image sequence is directly formulated as minimization of reprojection errors via the nonlinear least squares. Although several linear foromulations which ensure unique solution are also known, the nonlinear approach can be applied to much broader configurations, and gives more accurate results. We investigate fast optimization algorithm and several methods to avoid local minima for nonlinear reconstruction problem. We test our methods on recovering entire 3D shape from a building image sequence with several hundred images and several hundred feature points.
  • [Show abstract] [Hide abstract]
    ABSTRACT: 本論文では,建物のような屋外の大型の対象物を撮影した画像列から,対象物の全周形状を復元する問題を扱う.画像列からの形状復元は,一般に非線形最適化問題となるが,非線形最適化計算では,初期値設定により局所解に陥り,正解が得られないことが問題とされていた.それに対し,本論文では非線形最適化計算により直接解を求めるための,簡便な初期値設定法を提案する.この初期値設定は,全周復元の問題設定を反映したものであり,実場面において比較的広く適用できる.また,局所解に陥った場合に,解を修正することにより正しい形状復元を行う手法も提案する.最後に,画像数30〜90枚,特徴点数300〜1,200個規模のシミユレーションおよび実画像を用いた実験によって,本手法の有効性を示す. / We propose a method of setting the initial values of shape and camera parameters for structure from motion based on non-linear optimization method. The initial values of the parameters in the optimization process is crucial in solving the problem.. Thus the direct application of the non-linear optimization process has not been fully used in shape recovery. In this paper, first, we investigate an initialization method, which is very simple and relatively generic to the entire shape recovery problem. Then, we propose a method for restoring the optimization process from a local minimum. Experimental results using synthetic and real images have shown that our algorithm stably recover 3D shapes. 有
  • [Show abstract] [Hide abstract]
    ABSTRACT: 本論文では,屋外建物の全周形状復元問題に対して,簡便な初期値設定を行い,非線形最適化により直接解を求めることを提案する.画像列からの形状復元問題は,一般に非線形最適化問題となるが,非線形最適化では,初期値設定により局所解におちいり,正解が得られないことが問題とされていた.本論文で利用する初期値設定は,全周復元の問題設定を反映したものであり,実場面において比較的広く適用できる.また,局所解におちいった場合に,解を修正することにより正しい形状復元を行う手法も提案する.最後に,画像数数百枚,特徴点数数百点の規模のシミュレーションおよび実画像を用いた実験によって,本手法の有効性を示す. / We propose a procedure for recovering entire 3D shape from a building image sequence with non-linear optimization method. In this case, the initial values of the unknown variables in the optimization process is crucial. Thus the direct application of the non-linear optimization process has not been fully used in shape recovery. In this paper, first, we investigate an initialization method, which is very simple and relatively generic to the building shape recovery problem. Then, we propose a method for restoring the optimization process from a local minimum. Experimental results using real and synthetic image have shown that our algorithm stably recover 3D shapes.
  • [Show abstract] [Hide abstract]
    ABSTRACT: 画像列から物体形状とカメラ運動を同時に復元する運動からの形状復元は,非線形最適化問題として定式化できるが,安定化(局所解回避)と高速化(計算量低減)が大きな課題である.本論文では,非線形最適化アルゴリズムの1つである共役勾配法において,ヘッセ行列をブロック対角行列で近似した前処理行列を用いて計算量を低減する手法を提案する.シミュレーションおよび実画像の12種類のデータセットを用いた実験の結果,未知数が1,000個規模の問題に対して,従来のLevenberg-Marquardt法および共役勾配法の数倍から十数倍の高速化の効果を確認した. / Simultaneous recovery of shape and motion from image sequences is formulated as a nonlinear optimization problem. This paper proposes a fast algorithm named "block diagonal matrix preconditioned conjugate gradient method" characterized by block diagnalized approximation of Hessian. Experimental results using real and synthetic image data have shown that our algorithm reduces the calculation time by 80% to 95% compared with the Levenberg-Marquardt and conjugate gradient methods. 有
  • [Show abstract] [Hide abstract]
    ABSTRACT: 画像列中の各特徴点の2次元座標から特徴点の3次元位置と各画像を撮影したカメラの位置・姿勢を求める「運動からの形状復元」問題は非線形関数の最小化問題として定式化される.従来手法の多くは, 問題を簡略化するため, 実験室内の小物体や遠距離から撮影した建物の, 対象の片面などの一部を復元することを目的としていた.これに対して, 近距離から撮影した画像を用いて, 対象物の全周を復元することを考えると, 各画像には対象物の一部しか撮影されないため, ある特徴点を撮影している画像の数が少なくなり, 形状復元は困難な問題となる.本稿では, このような近距接撮影画像列からの建物の全周囲3次元復元を目的とし, このような状況で問題となる非線形関数の最小化における局所解を回避する方法を提案する.この手法により複数の建物を復元した結果を示し, 本手法の有効性を示す. / The problem of recovering 3 dimensional object shape shape from 2 dimensionl image sequence is called shape from motion problem. In the conventional approaches to the problem, primarily target scenes are limited to the in-door scenes, that is, a small object is taken from relatively far-distant cameras, because this situation significantly reduces difficulty of the problem. On the other hand, when we think of recovering entire building shape from images taken at the near disntance, the problem gets very difficult because each image contains only small part of the whole structure, which leads to the point that the feature point correspondence informations gets very poor. This paper describes a recovery method for this kind of situation. Experiments on real building recovery shows the efficiency of our methods.
  • [Show abstract] [Hide abstract]
    ABSTRACT: 数学定数である円周率πの多数桁の計算法として,級数展開を用いる方法とGauss-Legendreの公式などを用いる反復計算法が知られている.従来,N桁のπの値を得る計算量は,級数を用いるとO(N^2),反復計算法を用いるとO(N(log N)^2)とされ,Nが大きいときには反復計算法の方が有利であると考えられていた.本稿では,まず級数計算が2×2行列の積として表現できることを示し,次に隣接する2つの行列の積を再帰的に計算することによって級数を集約し,多倍長乗算を利用して級数の和をO(N(log N)^3)の計算量で求める方法について述べる.本手法を用いることによって,桁数Nが大きくなっても,級数による計算が反復計算法と同等の時間で行えることを確認した.また,3.2万桁から5.3億桁のπの計算時間の変化を調べた結果,級数の和を求めるChudnovskysの公式は,反復計算によるGauss-Legendreの公式よりも高速に計算できることが分かった. / A multiple-precision constant π of N digits is so far calculated by sum of series in O(N^2) computation time as well as by using iterational algorithm in O(N(log N)^2) time. In this paper, we first prove that the series for π calculation can be represented as a product of 2×2 matrices, and then propose a fast algorithm for calculating sum of series in O(N(log N)^3) time by recursively reducing the adjacent matrices. Using this algorithm, computation time of sum of series becomes comparable to that of iterational algorithm, and experimental results on calculating 32,000 to 530 million digits of π have shown that the sum of series using Chudnovskys formula is computed faster than the iterational algorithm using Gauss-Legendre algorithm. 有
  • [Show abstract] [Hide abstract]
    ABSTRACT: 多数桁の数学定数, 特にπや自然対数の計算法として簡単に導出できる級数展開を用いる方法と, πにおけるGauss-Legendreの公式等の反復計算法が知られている.πに関しては, 従来N桁の値を得る計算量は, 級数によるとO(N^2), 反復計算法によるとO(N(logN)^2)とされ, Nが大きい時には反復計算法の方が格段に有利であると言われていた.本稿ではある種の級数に対して, 隣接する級数の項を集約することにより, O(N(logN)^3)の計算量で級数の和を計算する計算法を示した.この方法によって桁数Nが大きい時にも, 従来計算時間的に反復計算法より不利とされてきた級数による計算が, 同等の時間で行える, 本手法を用いることにより, 3.2万桁から5.3億桁のπの計算に関して, 級数の和を用いたChudnovskyの公式を, 反復計算によるGauss-Legendreの公式よりも高速に計算できることが明らかになった. / Multiple-precision mathematical constants, such as π or e are known to be calculated by sum of series. On the other hand, much faster calculation method that use iteration are known for some constants such as π. For the case of π, N digits calculation time by method of sum of series is said to be O(N^2), and that of iterational method is O(N(logN)^2).Thus, for large N, iterational method is far more efficient than that of sum of series. In this paper, we propose a fast algorithm of calculating sum of series in O(N(logN)^3)time by recursively reducing adjacent terms of series. With this algorithm, calculation time of sum of series become comparable to that of iterational method in case of large N. Experimental results on calculating 32, 000 to 530 million digits of π showed that the Chudnovsky formula which uses sum of series can be calculated faster than the Gauss-Legendre method which uses iterational method.
  • [Show abstract] [Hide abstract]
    ABSTRACT: 従来のコンピュータビジョン研究では, 「観察者」はシーンの外に位置し, シーンに触れない, あるいはシーンを変化させないという条件の下で, 三次元情報を正確に取得することを行っていた.これに対し, 我々は, ビジョンによる3次元情報復元というタスクと, その情報に基づいて物体に作用するというアクションタスクを融合し, 能動的にシート中の物体を移動するなど, シートの改変を行うことによって物体の形状および力学的特性を安定に取得する能動認識(Active Recognition)の研究を進めている.本研究では, 能動認識の試みとして, 物体を押すことによって生じる移動を観測した視覚情報から, 物体の力学的特性である重心位置の推定を行った.机の上に置かれたトレイを, ロボットアームにより移動させ, 重心位置を推定する実験を行った結果, 重心位置の場所にかかわらず, 安定に重心推定が可能であることがわかった. / The task of computer vision is to recover accurate 3-dimensional shape of objects from 2-dimensional images. In such a vision task, the observer is assumed to be outside of the scene, and this means that the scene configuration are never destroyed by the observre. In this paper, we propose a new paradigm "active recognition", in which the object information is obtained by cooperation of visual and physical systems which can reform the scene configuration. Experiments on the estimation of the center of gravity was performed with a robot arm and a camera that were used to push and track the object. The stable results have shown the effectiveness of the integration of visual and physical information.