Bruce Merry

Animation space: A truly linear framework for character animation

Skeletal subspace deformation (SSD), a simple method of character animation used in many applications, has several shortcomings; the best-known being that joints tend to collapse when bent. We present animation space, a generalization of SSD that greatly reduces these effects and effectively eliminates them for joints that do not have an unusually large range of motion.While other, more expensive generalizations exist, ours is unique in expressing the animation process as a simple linear transformation of the input coordinates. We show that linearity can be used to derive a measure of average distance (across the space of poses), and apply this to improving parametrizations.Linearity also makes it possible to fit a model to a set of examples using least-squares methods. The extra generality in animation space allows for a good fit to realistic data, and overfitting can be controlled to allow fitted models to generalize to new poses. Despite the extra vertex attributes, it is possible to render these animation-space models in hardware with no loss of performance relative to SSD.

Compression of dense and regular point clouds

We present a simple technique for single-rate compression of point clouds sampled from a surface, based on a spanning tree of the points. Unlike previous methods, we predict future vertices using both a linear predictor, which uses the previous edge as a predictor for the current edge, and lateral predictors that rotate the previous edge 90° left or right about an estimated normal.By careful construction of the spanning tree and choice of prediction rules, our method improves upon existing compression rates when applied to regularly sampled point sets, such as those produced by laser range scanning or uniform tesselation of higher-order surfaces. For less regular sets of points, the compression rate is still generally within 1.5 bits per point of other compression algorithms.

A comparison of linear skinning techniques for character animation

Character animation is the task of moving a complex, artificial character in a life-like manner. A widely used method for character animation involves embedding a simple skeleton within a character model and then animating the character by moving the underlying skeleton. The characters skin is required to move and deform along with the skeleton. Research into this problem has resulted in a number of skinning frameworks. There has, however, been no objective attempt to compare these methods. We compare three linear skinning frameworks that are computationally efficient enough to be used for real-time animation: Skeletal Subspace Deformation, Animation Space and Multi-Weight Enveloping. These create a correspondence between the points on a characters skin and the underlying skeleton by means of a number of weights, with more weights providing greater flexibility. The quality of each of the three frameworks is tested by generating the skins for a number of poses for which the ideal skin is known. These generated skin meshes are then compared to the ideal skins using various mesh comparison techniques and human studies are used to determine the effect of any temporal artefacts introduced. We found that Skeletal Subspace Deformation lacks flexibility while Multi-Weight Enveloping is prone to overfitting. Animation Space consistently outperforms the other two frameworks.

Parallel, Realistic and Controllable Terrain Synthesis

The challenge in terrain synthesis for virtual environments is to provide a combination of precise user control over landscape form, with interactive response and visually realistic results.

Compression of Dense and Regular Point Clouds

We present a simple technique for single‐rate compression of point clouds sampled from a surface, based on a spanning tree of the points. Unlike previous methods, we predict future vertices using both a linear predictor, which uses the previous edge as a predictor for the current edge, and lateral predictors that rotate the previous edge 90°left or right about an estimated normal.

Moving Least-Squares Reconstruction of Large Models with GPUs

Modern laser range scanning campaigns produce extremely large point clouds, and reconstructing a triangulated surface thus requires both out-of-core techniques and significant computational power. We present a GPU-accelerated implementation of the moving least-squares (MLS) surface reconstruction technique. We believe this to be the first GPU-accelerated, out-of-core implementation of surface reconstruction that is suitable for laser range-scanned data. While several previous out-of-core approaches use a sweep-plane approach, we subdivide the space into cubic regions that are processed independently. This independence allows the algorithm to be parallelized using multiple GPUs, either in a single machine or a cluster. It also allows data sets with billions of point samples to be processed on a standard desktop PC. We show that our implementation is an order of magnitude faster than a CPU-based implementation when using a single GPU, and scales well to 8 GPUs.

Accelerating kd-tree Searches for all k-nearest Neighbours

Finding the k nearest neighbours of each point in a point cloud forms an integral part of many point-cloud processing tasks. One common approach is to build a kd-tree over the points and then iteratively query the k nearest neighbors of each point. We introduce a simple modification to these queries to exploit the coherence between successive points; no changes are required to the kd-tree data structure. The path from the root to the appropriate leaf is updated incrementally, and backtracking is done bottom-up. We show that this can reduce the time to compute the neighbourhood graph of a 3D point cloud by over 10%, and by up to 24% when k = 1. The gains scale with the depth of the kd-tree, and the method is suitable for parallel implementation.

Normal transformations for articulated models

It is well-established that when a matrix is used to transform a rigid object, the normals should be transformed by the inverse transpose of that matrix. For skeletally animated models, it is common to apply this approach to the blended matrix that animates each vertex. This is only an approximation, as it assumes that the blended matrix is locally constant. We derive a correct method for normal transformation in skeletally animated models, and examine the errors introduced by two approximations.

Analytic simplification of animated characters

Traditionally, levels of detail (LOD) for animated characters are computed from a single pose. Later techniques refined this approach by considering a set of sample poses and evaluating a more representative error metric. A recent approach to the character animation problem, animation space, provides a framework for measuring error analytically. The work presented here uses the animation-space framework to derive two new techniques to improve the quality of LOD approximations. Firstly, we use an animation-space distance metric within a progressive mesh-based LOD scheme, giving results that are reasonable across a range of poses, without requiring that the pose space be sampled. Secondly, we simplify individual vertices by reducing the number of bones that influence them, using a constrained least-squares optimisation. This influence simplification is combined with the progressive mesh to form a single stream of simplifications. Influence simplification reduces the geometric error by up to an order of magnitude, and allows models to be simplified further than is possible with only a progressive mesh. Quantitative (geometric error metrics) and qualititative (user perceptual) experiements confirm that these new extensions provide significant improvements in quality over traditional, naïve simplification; and while there is naturally some impact on the speed of the off-line simplification process, it is not prohibitive.

Fast in-place binning of laser range-scanned point sets

Laser range scanning is commonly used in cultural heritage to create digital models of real-world artefacts. A large scanning campaign can produce billions of point samples—too many to be manipulated in memory on most computers. It is thus necessary to spatially partition the data so that it can be processed in bins or slices. We introduce a novel compression mechanism that exploits spatial coherence in the data to allow the bins to be computed with only 1.01 bytes of I/O traffic for each byte of input, compared to 2 or more for previous schemes. Additionally, the bins are loaded from the original files for processing rather than from a sorted copy, thus minimizing disk space requirements. We demonstrate that our method yields performance improvements in a typical point-processing task, while also using little memory and guaranteeing an upper bound on the number of samples held in-core.