Action Synopsis: Pose Selection and Illustration 1. This paper introduced a method that produces an action synopsis for presenting motion in still images. 2. Key poses are first selected from each motion, and then rendered in still images. 3. Key points: 3.1 Extracting motion aspects Given a sequence of skeletal poses, a small number of motion aspects is calculated. A motion aspect is an attribute of the motion, by which we define inter-pose distances. In the current implementation, four aspects, namely, joint positions, angles, speed and angular velocity, are used. 3.2 Dimensionality reduction For each motion aspect, an affinity matrix, which encapsulates weighted distances among all poses, are calculated. The dimension of the motion curve is reduced by analyzing the affinity matrices with a Replicated Multi-dimensional Scaling (RMDS) [McGee 1978]. The technique defines a reduced dimensional space (typically of 5-9 dimensions) in which the salient features of the various motion aspects are kept. 3.3 Pose selection The local extreme points along the motion curve are located, which are associated with the extreme poses of the motion. This stage generates a hierarchy of prioritized poses. 3.4 Synopsis illustration the frames associated with the selected poses are composed into an image. The selected frames can either be presented side by side, or be composed into a single image. To further enhance the image, some instances can be rendered semi-transparently, thus reducing the cluttering of the resulting image and highlighting the more significant poses. 4. Potential improvments 4.1 A concise key pose selection is desired, especially for cyclic motions. 4.2 Better camera settings are desired. //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// Representing motion in a static image: constraints and parallels in art, science, and popular culture 1. This paper illustrated and discussed five different types of representations of motion. 2. Five ways to represent motion: (1) dynamic balance, broken symmetry, even instability. (2) multiple images. (3) affine shear. (4) blur. (5) vector-like lines superimposed on an image. 3. Four criteria (1) evocativeness (2) the clarity of the object represented (3) the direction of depicted motion (4) precision 4. Evaluating the 5 motion representations (1) dynamic balance: good at evocativeness, good at clarity of object, fair good at motion direction, bad at motion precision (2) multiple images: good at evocativeness, good at clarity of object, bad at motion direction, good at motion precision (3) affine shear: good at evocativeness, good at clarity of object, good at motion direction, bad at motion precision (4) photographic blur: good at evocativeness, bad at clarity of object, bad at motion direction, bad at motion precision (5) action line: good at evocativenss, good at clarity of object, good at motion direction, good at motion precision //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// 3D motion retrieval with motion index tree 1. This paper presented a novel content-based 3D motion retrieval algorithm. 2. Challenges 2.1 An effective motion representation is needed. 2.2 An efficient match algorithm is demanded to calculate the similarity between two motions. 2.3 A key-frame extraction algorithm is required. 3. Key points 3.1 Based on a hierarchical motion description, the motion library is partitioned and a motion index tree is constructed using the nearest neighbor rule based dynamic clustering algorithm. 3.2 The motion index tree serves as a classifier to determine the sub-library that contains the promising similar motions to the query sample. 3.3 The similarity between two motions is calculated through elastic match. 3.4 An adaptive clustering-based key-frame extraction algorithm is adopted to improve the efficiency of the similarity calculation. 4. Potential improvements 4.1 Motion retrieval with sketched key frames as a query 4.2 Better pose similarity metrics //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// //////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////// A System for Analyzing and Indexing Human-Motion Databases 1. This paper demonstrated a data-driven approach for representing,compressing, and indexing human-motion databases. 2. The basic approach is to build an implicit model of every distinct body pose seen in a motion database, and to cluster these poses into groups that can be effectively interpolated using simple linear models. Motion sequences can then be classified and indexed by the trajectories that they take through the set of pose clusters. The linear models also significantly reduce the storage requirements of the database. 3. Key points 3.1 Principal marker selection PFA algorithm is used to determine the subset of database features used to construct the local linear model classifier. 3.2 Piecewise-linear modeling A piecewise-linear model is constructed using a divisive clustering approach that, at each level, attempts to obtain a best-fit linear model of a user-specified dimension, d, and error tolerance,ŠĆ. This best-fit model is constructed via successive applications of PCA. 4. Applications 4.1 Motion compression 4.2 Estimating Human Motions from a Reduced Marker Set 4.3 Motion database indexing