• No se han encontrado resultados

CAPÍTULO IV. - ​APLICACIÓN DEL MODELO EN EL ÁREA DE CONSERVACIÓN Y USO SOSTENIBLE YUNGUILLA

4.6. Cuantificación de indicadores

Realistic facial animation remains as a fundamental challenge in computer graphics. Since the pioneering work of Parke [24], a large body of literature on modeling and animating faces has been published in the last four decades. A good overview can be found in the textbook by Parke and Waters [23] and in the survey by Noh and Neumann [21]. In the context of this paper, we focus on publications that address performance-driven facial ani- mation and muscle-based face modeling.

Remarkably, one of the oldest publications in this context is the one that uses three- dimensional sparse motion capture marker data to control facial movement of computer- generated models [32]. The system synthesizes expressions by changing texture coordinates calculated from the positions of the markers on the performer’s face. Eisert and Girod [9] model a face with a B-spline surface, and analyze facial expressions into feature point positions to estimate the facial animation parameters of the MPEG-4 standard. Guenter [13] capture both 3D geometry and shading information of a human face, and reproduce photorealistic expressions. In all of these methods, the locations of the markers are used to drive the 3D model. Since the markers usually are quite sparse compared to the dense surface mesh of the model, an interpolation function is typically used to deformed the mesh so that vertices in between the markers are displaced properly.

Another category of performance-driven animation is to synthesize expressions by blending pre-modeled key expressions. The animation is achieved by computing a set of

blending weights that minimize the Euclidean distance between the corresponding markers on the actor’s face and the 3D model. Pighin et al. [27] reconstruct the geometry and tex- ture of an individual face from several face images taken from different viewangles. They also model basic expressions and generate novel expressions by blending them. Later, they propose a method to find the blending weights by minimizing an error function over the set of pre-modeled expressions and face positions spanned by the model [26]. Kouadio et al. [17] animate a synthetic character by extracting the interpolation weights from the feature points traced by an optical capture system. Choe and Ko [5] develop an artist-in-the-loop method for analyzing captured expressions. The expressions are synthesized by a linear combination of the elements in a muscle actuation basis which consists of face shapes re- sulting from the contraction of each single facial muscle. Typically, the basis elements need to be resculpted a number of times to obtain satisfactory results. In the approach proposed by Chuang and Bregler [6], the 2D key expressions are automatically found from the track- ing data, and the corresponding 3D key shapes of a face model are created manually. Facial animation is produced by applying the blending weights recovered from facial feature de- composition to the 3D key shapes. However, a complete bank of 3D key shapes must be built for any new subject, which is a tedious task.

Some approaches involve mapping the motion of a facial expression from the source model to the target model directly [20]. Since the target model may have different shape, the source motion vectors need to be transformed to follow the curvature of the new face shape. In order to generate delicate skin deformations, dense mesh motion is required as input. But this may not be available from some motion capture systems. Moreover, dense correspondences between the source and target models should be established for motion retargeting, which is difficult if the source and target shapes are very different.

In facial animation, desire for improved realism has driven researchers to extend ge- ometric models with physical models of facial anatomy which attempt to emulate the in- fluence of muscle contraction onto the skin surface by approximating the biomechanical properties of skin [15, 16, 18, 25, 30, 33]. Animating a detailed muscle-based model can be rather difficult since facial muscles contract in a complex coordinated manner to generate expressions. A solution to this problem is to automatically determine muscle activations from the facial motion data. Terzopoulos and Waters [31] extract muscle contraction pa- rameters based on the position of facial features tracked by snakes. Morishima et al. [19] use 2D marker positions as input for a neural network which estimates muscle actuation parameters. Both of these approaches require heavy makeup of the actor’s face. Some techniques compute an optical flow from the video sequence and decompose the flow into muscle activations. Essa et al. [12] use a physical face model and develop a system to estimate muscle contractions that match optical flow input based on feedback control the- ory. Decarlo and Metaxas [8] employed a similar model that incorporates variations in head shape using anthropometric measurements. In [1, 4], a 2D quasi-static finite element model is used to simulate movements of the lips [1] or the facial skin surface [4]. Given marker data, the authors use a steepest descent iterative solver to calculate the lip model parameters or the facial muscle activations that best track the motion data. More recently, Sifakis et al. [29] employ an optimization framework to determine muscle activations that track a sparse set of surface landmarks, and used it for speech animation [28]. However, the computa- tional complexity of the nonlinear optimization process makes this method unsuitable to

retarget facial expressions in real-time.