SPC-Net: A New Scalable Point Cloud Compression Framework for Both Machine and Human Vision Tasks

Lei Liu, Zhihao Hu, Dong Xu

Keywords: point cloud, compression, scalable coding

Abstract: Recently, point cloud process and analysis have attracted increasing attention in various machine vision tasks. Therefore, some point cloud compression algorithms are developed. However, such compression algorithms are developed for human vision while most of the point cloud data will be used for automated point cloud analysis (e.g., detection of abnormal event and early warning in autonomous driving) and may not be seen by humans. To this end, we design a new scalable point cloud compression framework (SPC-Net) for both machine and human vision tasks, in which a scalable bit-stream will be used to describe the point cloud for both machine vision and human vision tasks. For machine vision tasks, only part of the bit-stream will be transmitted for bit-rate saving, while the full bit-stream will be transmitted when used for the human vision task. Additionally, we propose a new octree depth level predictor to automatically predict the optimal depth level in order to control the bit-rate cost for the machine vision tasks.As a result, for simple objects/scenarios, we will use fewer depth levels with less bits for the machine tasks, while for complex objects/scenarios, we prefer deeper depth levels of octree with more bits for machine tasks comprehensive. Experimental results on different datasets (e.g., ModelNet10, ModelNet40, ShapeNet and ScanNet) demonstrate that our proposed scalable point could compression framework SPC-Net achieves better performance on the machine vision tasks (e.g., classification, segmentation and detection) without degrading the performance of the human vision task.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)