Compression without Compromise: Optimizing Point Cloud Object Detection with Bottleneck Architectures For Split Computing

• Published in: 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW) pp. 1 - 6
• Main Authors: Kashyap, Vinay, Ahuja, Nilesh, Tickoo, Omesh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.07.2024
• Subjects: Accuracy; Bit rate; Compression; Computer architecture; Machine Vision; Object detection; Performance evaluation; Point cloud compression; Point Clouds; Split Computing; Transforms
• ISSN: 2995-1429
• DOI: 10.1109/ICMEW63481.2024.10645461

Recent advances in data capture technologies have brought point cloud data to the forefront for many perception tasks in robotics, autonomous driving, etc. However deep learning inference on point cloud data requires significant computing capability, which would place extensive load on resource constrained low-power edge devices. To address these computing challenges, there has been a large push towards performing analytics on the cloud by compressing and transmitting the captured data. Point cloud compression is challenging due to the required computation, the fact that the data is inherently unordered, and sensitivity to network factors such as bitrate. This can cause losses and artifacts during decompression, leading to drops in the performance of downstream analytics tasks. Building on the premise of the client-edge co-inference, where compute is split between the client device and the edge server, we propose to reduce bitrate without sacrificing inference accuracy by using a task specific split-DNN approach. Low complexity neural networks called 'bottleneck units' are introduced at the split point to transform the intermediate layer features into a lower-dimensional representation that are better suited for compression and transmission.