Environmental perception is widely recognized as the crucial part of intelligent transportation systems, in which point cloud semantic segmentation is an important task. High-performance semantic segmentation can help the system better understand the 3D world. This paper proposes a strategy of stacking base networks to construct high-performance segmentation networks. To reduce the model size and computational resource consumption, a novel lightweight method is proposed for compressing high-performance stacked point cloud segmentation networks. The proposed framework, called Stacked and Distillation Network Framework (SDNF), involves transferring the dark knowledge of the teacher network to the student network through the point cloud's structural information, attention information, and predicted labels. Finally, experiments with S3DIS dataset show that the proposed SDNF significantly improves the segmentation performance of both CNN/MLP-based and Transformer-based networks while keeping the model size minimal.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    Stacked and Distillation Network Framework for Point Cloud Semantic Segmentation


    Contributors:
    Han, Jiawei (author) / Liu, Kaiqi (author)


    Publication date :

    2023-09-24


    Size :

    3208870 byte





    Type of media :

    Conference paper


    Type of material :

    Electronic Resource


    Language :

    English



    Deep Learning on 3D Point Cloud for Semantic Segmentation

    Ning, Zhihan / Tang, Linlin / Qi, Shuhan et al. | Springer Verlag | 2021



    Deep Learning on 3D Point Cloud for Semantic Segmentation

    Ning, Zhihan / Tang, Linlin / Qi, Shuhan et al. | TIBKAT | 2022


    SAT3D: Slot Attention Transformer for 3D Point Cloud Semantic Segmentation

    Ibrahim, Muhammad / Akhtar, Naveed / Anwar, Saeed et al. | IEEE | 2023


    TransKD: Transformer Knowledge Distillation for Efficient Semantic Segmentation

    Liu, Ruiping / Yang, Kailun / Roitberg, Alina et al. | IEEE | 2024