FMP-Net: Fractal Multi-Gate Mixture-of-Experts Panoramic Segmentation for Point Cloud
- Resource Type
- Conference
- Authors
- Yang, Chengzhe
- Source
- 2022 IEEE International Conference on Multimedia and Expo (ICME) Multimedia and Expo (ICME), 2022 IEEE International Conference on. :1-6 Jul, 2022
- Subject
- Communication, Networking and Broadcast Technologies
Components, Circuits, Devices and Systems
Computing and Processing
Signal Processing and Analysis
Point cloud compression
Laser radar
Filtering
Semantics
Logic gates
Feature extraction
Fractals
Point Cloud
Panoramic Segmentation
MMoE
Binary Tree
- Language
- ISSN
- 1945-788X
Panoramic segmentation in LiDAR point clouds is in its nascent stage, which unifies the semantic and instance seg-mentation tasks that receive increasing attentions. This paper proposed a fractal multi-gate mixture-of-experts panoramic segmentation network (FMP-Net) consists of encode and de-code, which is able to perform panoramic segmentation tasks on large-scale scenes point clouds. The fractal multi-gate mixture-of-experts (FMMoE) module is proposed in the de-code part, which integrates the multi-gate mixture-of-experts (MMoE) module and binary tree. The FMMoE module is able to divide the features extracted by the encoder step by step, eventually determining the features to be used for each sub-task in panoramic segmentation. We used a hierarchical clus-tering algorithm to obtain instance information while improving the overall network inference speed. The experimental results illustrated that our proposed method achieve Panoptic Quality(PQ) of 40.7% and the Intersection over Union(mIoU) of 62.3% on the public dataset NuScenes. The source code is available at: https://github.com/ChazYang/FMP-Net.