Cross-layer Features Fusion Network with Attention MLP for ALS Point Cloud Segmentation

In recent years, many semantic segmentation algorithms for airborne laser scanning (ALS) point clouds based on the encoding-decoding structure have been proposed, but the influence of different layers of feature fusion methods in the decoding process on the final segmentation effect has not been considered. In this study, we propose a cross-layer features fusion network with attention MLP (CFFA-Net) for ALS point cloud segmentation by investigating how features from different layers are fused in the decoding process. In detail, by designing an attention MLP (AMP) module in the encoding process, we ensure that the down-sampling process of the network reduces the loss of crucial spatial information. For the different layers of features in the decoding process, the cross-layer features fusion (CFF) module performs different combinations of fusion experiments and finally obtains the best fusion method. With the AMP module and the CFF module, the CFFA-net can be better adapted to the segmentation of the ALS point cloud with complex backgrounds, sparse point distributions, and large variations in object scales. Compared with the baseline, our method shows superior segmentation results on both the Vaihingen 3D dataset and the DFC2019 dataset.