Automatic Hand-Over Animation using Principle Component Analysis Nkenge Wheatland Sophie J ¨ org Victor Zordan UC Riverside * Clemson University UC Riverside Abstract This paper introduces a method for producing high quality hand motion using a small number of markers. The proposed “hand- over” animation technique constructs joint angle trajectories with the help of a reference database. Utilizing principle component analysis (PCA) applied to the database, the system automatically determines the sparse marker set to record. Further, to produce hand animation, PCA is used along with a locally weighted regression (LWR) model to reconstruct joint angles. The resulting animation is a full-resolution hand which reflects the original motion without the need for capturing a full marker set. Comparing the technique to other methods reveals improvement over the state of the art in terms of the marker set selection. In addition, the results highlight the ability to generalize the motion synthesized, both by extending the use of a single reference database to new motions, and from distinct reference datasets, over a variety of freehand motions. Keywords: character animation, motion capture, hand motion, dimensionality reduction, PCA 1 Introduction Producing quality whole-body motion involves the movement of the hand in relation to the rest of the body. However, using a motion capture system, it can be difficult to record the full body of a moving person while also capturing the hand and all of its detail because the whole-body and hand appear at largely different scales. While it is possible to record a high-resolution capture of the hand through a comprehensive set of markers (typically 13-20 markers), this is often only possible in a small capture region, isolating the motion of the hand. However, in a larger, full-body capture region, the complete set of markers becomes difficult to discern, and so this * e-mail:wheatlan@cs.ucr.edu approach is usually abandoned in lieu of the capture of a smaller set of markers (2-6 markers) coupled with a “hand-over” process for reconstructing the full hand animation [Kang et al. 2012]. In this paper, we propose a robust technique to accomplish the latter that both automatically selects the “sparse” marker set to record, and subsequently produces joint trajectories for a full hand from the sparse marker set. Our technique employs a combination of principle component anal- ysis (PCA) [Bishop 1995] to construct a low-dimensional repre- sentation of the hand data along with linearly weighted regression (LWR) [Atkeson et al. 1997] to aid in the reconstruction. Starting from a reference database that is recorded using a full-resolution marker set, we first determine the best sparse marker set to record based on the PCA representation of the data. We experiment with different test sizes for the marker set to record, specifically reduced marker sets of six and three markers, and we compare our selec- tion method with different ones proposed for selecting the mark- ers, including manual selection, following Hoyet et al. [2012], and a method that uses representative cluster-based search for selec- tion [Kang et al. 2012]. In contrast, the technique in this paper com- putes the marker set directly from the PCA, and our findings show that this marker set is superior to the other methods of selection for the reconstruction technique we propose. For reconstruction, our method employs a second PCA in a synthesis step combined with LWR. Starting from a test query that records only the sparse marker set, we use LWR to build a locally sensitive model between the markers and the principle components. We use American Sign Language (ASL) as our primary testbed. ASL is an important and interesting freehand application of hand motion. Further, it includes a rich, diverse set of configuration poses for the hands. We show that we can construct new (unseen) ASL signs with high-visual quality using a simple, generic ASL database. Generalization of the database reveals that we can use our technique to capture other motions, such as counting. Our effort holds close similarities to previous work, especially the full-body motion control of Chai and Hodgins [2005]. In contrast, our main contributions include the distinct exploration of rich hand data, such as ASL, as well as our method for determining the best reduced marker set to take advantage of the power of dimensionality reduc- tion realized by PCA. Further, our approach is far simpler and lends itself to ease-of-use and re-implementation. Our approach also has notable advantages over other related papers for hand-over anima- tion, such as the work of Hoyet et al. [2012] and Kang et al. [2012] in that we compute the best reduced marker set directly, rather than selecting it manually or through brute-force search. Compared to other techniques, ours is both simple to implement and fast to compute, striking a valuable compromise which is likely to lead to greater adoption for commercial use. 2 Related work The detailed and subtle motions of the hands are hard to cap- ture. Several approaches for recording have been suggested, each with advantages and disadvantages. In particular, optical motion capture systems, while being very accurate, can require substan- tial post-processing to handle occlusions and mislabelings. Cyber Gloves [2013] and the like are robust to captures in larger spaces, but they require regular calibration [Wang and Neff 2013] and do