In computer graphics, triangle meshes are ubiquitous as a representation of surface models. Processing of this kind of data, such as compression or watermarking, often involves an unwanted distortion of the surface geometry. Advanced processing algorithms are continuously being proposed, aiming at improving performance (compression ratio, watermark robustness and capacity), while minimizing the introduced distortion. In most cases, the final resulting mesh is intended to be viewed by a human observer, and it is therefore necessary to minimise the amount of distortion perceived by the human visual system. However, only recently there have been studies published on subjective experiments in this field, showing that previously used objective error measures exhibit rather poor correlation with the results of subjective experiments. In this paper, we present results of our own large subjective testing aimed at human perception of triangle mesh distortion. We provide an independent confirmation of the previous result by Lavoué et al. that most current metrics perform poorly, with the exception of the MSDM/MSDM2 metrics. We propose a novel metric based on measuring the distortion of dihedral angles, which provides even higher correlation with the results of our experiments and experiments performed by other researchers. Our metric is about two orders of magnitude faster than MSDM/MSDM2, which makes it much more suitable for usage in iterative optimisation algorithms.

No abstract

Mesh segmentation (clustering) is a useful tool, which improves compression performance. On the other hand, per-partes processing of meshes often leads to new types of artifacts-cracks and shifts on the borders between clusters. These artifacts are detected by both, Human Visual System (HVS) and perceptually-motivated distortion metrics. In this paper, we present a post processing algorithm, which aims at reducing such artifacts without needing any additional data-using only information about the cluster distribution that is already present at the decoder. A rigid transformation, which minimises the border artifacts, is iteratively computed and applied per cluster. Our experiments show that this approach leads to a reduction of distortion, as measured by the STED metric, by up to 18% for low bitrates. We also present visual results confirming that the improvement is well visible.