| 초록 |
With the increased interest in point cloud processing, point cloud sampling techniques have also been attracting attention. As more models can calculate point clouds directly, working time has become an essential factor. The down-sampling process can solve this problem. Existing sampling algorithms have performed constant sampling methods regardless of the characteristics of task-model learning. However, this weakness has limitations in improving the performance of the task model in learning, and such task-agnostic methods perform too low when the sampling rate is high. Therefore, this paper proposes a novel down-sampling model network based on deep learning tasks. The proposed network utilizes fully connected layers to extract meaningful features from input sequences, and adds positional encoding instead of a conventional convolution concept. By introducing positional encoding into down sampling, the proposed network learns about the relationship between point clouds, and generates a task-oriented sampling methodology. Furthermore, the network incorporates skip connections to preserve important information during the down-sampling process. The proposed model outperforms several state-of-the-art models in terms of classification accuracy. With the fast inference time of the proposed network, it can be used in various applications, and our approach provides a promising solution for down-sampling tasks in different point cloud applications. |
