Multi-Modal Evaluation of 3D Point Clouds Images: A Novel No-Reference Approach Using a Multi-Stream Attentive Architecture

Most existing Point Cloud Quality Assessment (PCQA) methods do not consider the local structures among points, which can impact the overall perceived quality. In this paper, we introduce a novel and efficient no-reference objective metric for PCQA that takes into account the intrinsic feature affinities of points using a fully attention-based network, which results in extracting relevant information to the local structures of the 3D content. In addition, we employ information from two modalities: a suitable 2D projection of the PC and a relevant subset of the native 3D point cloud data. The rationale is that each modality may be more sensitive to different distortion types and thus contribute to the overall quality assessment. To evaluate the performance of our method, we conducted experiments on well-known 3D Point Clouds Quality Assessment benchmarks for PC compression. Our results demonstrate that our multi-modal attention-based PCQA metric is competitive with state-of-the-art methods in terms of both effectiveness and reliability. In particular, our method is able to capture local structures and provide more accurate quality assessments, even better than most full-reference metrics, with a moderate computational cost.