The alpha parallelogram predictor: A lossless compression method for motion capture data

Abstract

Motion capture data in an uncompressed form can be expensive to store, and slow to load and transmit. Current compression methods for motion capture data are primarily lossy and cause distortions in the motion data. In this paper, we present a lossless compression algorithm for motion capture data. First, we propose a novel Alpha Parallelogram Predictor (APP) to estimate the DOF (degree of freedom) of each child joint from those of its immediate neighbors and parents that have already been processed. The prediction parameter of the predictor, which is referred to as the alpha parameter, is adaptively chosen from a carefully designed lookup table. Second, we divide the predicted and actual values into three components: sign, exponent and mantissa. We then compress their corrections separately with context-based arithmetic coding. Compared with other lossless compression methods, our approach can achieve a higher compression ratio with a comparable compression time. It can be used in situations where lossy compression is not preferred.

Introduction

In recent years, motion capture data are widely used in many applications, in particular in movies and games. The advance in data-driven animation technologies rapidly produces huge collections of motion capture data, and it becomes necessary to design effective compression algorithms for the storage and transmission of these data. There are a number of compression methods proposed for motion capture data [1], [2], [6], [10], [23], [27]. Although most of these compression methods have shown to produce good compression performances, they are all based on lossy compression and hence modify the original data. Scientists and engineers typically do not prefer the idea of having their data modified by a process beyond their control and therefore often refrain from using these lossy compression algorithms. In addition, any lossy compression methods for motion capture data, whether they are orientation-based or position-based, introduce errors to the motion data, causing various perceptual artifacts. For example, the well-known foot-skating artifact [18] is caused by the failure in preserving clean footprints. Although we may adopt inverse kinematics to reduce or minimize this error [1], [2], [27], the residual error can still degrade the visual quality of the motion data. Furthermore, if we want to edit/modify a motion file compressed with a lossy compression method, we will need to uncompress it before editing and then compress it again afterwards. Each time we compress the motion file, the lossy compression method will introduce additional errors to the file. Hence, a motion file needed to go through a number of such editing operations at different times will result in some compound errors.

To address these problems, we propose in this paper a lossless compression method for motion capture data. To our knowledge, there has not been any work published that addresses the lossless compression of motion capture data. Our method can be summarized as follows. We first represent the motion capture data as a matrix, with all the elements split into several equal-sized clusters. For each cluster, we propose a novel Alpha Parallelogram Predictor (APP) to estimate a DOF (degree of freedom) value for each child joint from those of its immediate neighbors and parents that have already been processed. Although the APP is an extension of the parallelogram predictor for mesh compression [15], [26], it is significantly different from the parallelogram predictor due to the introduction of a prediction parameter, referred to as the alpha parameter, and the indexing storage of alpha. Through the introduction of this alpha parameter, the APP can obtain more accurate prediction results and smaller correction values than those of the parallelogram predictor. As a result, it produces a better compression ratio. To address the storage overhead of the alpha values, we have carefully designed a lookup table from which the prediction parameter is chosen. In this way, we only need to store the indices to the lookup table to save storage space. Finally, after the prediction step, both the predicted and the actual values are separated into three components: sign, exponent and mantissa. The correction of each component is compressed by means of arithmetic coding. Our experiments show that the proposed method outperforms the lossless MPEG-FBA compression method and two general compression tools, gzip and Rar, in terms of compression ratio. The proposed method is completely lossless and nicely complements the lossy compression methods for motion capture data. The contributions of our work can be summarized as follows:

• We introduce an alpha parameter to the traditional parallelogram predictor [15], [26] to improve the prediction accuracy and hence the compression ratio.

• To reduce the storage overhead of the alpha values, we propose to select the alpha values from a carefully designed lookup table.

• We have adapted the lossless float-point compression algorithm [16] to post-process the correction values obtained from the APP. As the APP can work better than the traditional parallelogram predictor, different context policies can be used for the actual exponent and mantissa corrections, based on the most frequently called sequence in Algorithm 1.

The remainder of this paper is organized as follows. Section 2 provides a review of related works. Section 3 describes the splitting of the motion capture data and the proposed APP, while Section 4 describes the floating-point compression method, which deals with the actual and predicted values produced by the APP. Section 5 presents some experimental results and evaluates the proposed method. Finally, Section 6 gives the conclusion remarks and discusses possible future works.