PG_Net details


Short Name PG_Net
Long Name FG_Net (Feature Geometric Net)
Description This work presents a general deep learning framework for large-scale point clouds processing without voxelizations, called PGNet, which achieves accurate recognition of point clouds with real-time processing. We design noise removal methods as preprocessing which enhances the performance of subsequent tasks including classification and segmentation. We propose a novel point clouds down-sampling strategy combining inverse density sampling and gumbel softmax sampling, which strikes a great balance between efficiency and performance for large-scale point clouds processing. Through elaborate design combining feature level correlation mining and deformable convolutions based geometric aware modeling, the local feature relationships and geometric patterns can be captured. The attention mechanism is also adopted to enhance the global non-local long-range feature correlations. Finally, the feature pyramid based residual learning network architecture is designed to combine patterns at different resolutions in a memory-efficient way. The effectiveness of our method has been validated extensively on various large-scale and real-scene indoor and outdoor datasets. It turns out that it gives better performance in point clouds classification and segmentation compared with state-of-the-art methods while guaranteeing real-time performance. We have also done weakly supervised transfer learning to demonstrate the generalization capacity of our methods. Some visualizations of the inner activation of network modules are presented to interpret what has been learned by our framework.
Reference Kangcheng Liu, and Ben M. Chen. (FG-Net: Fast Large-Scale LiDAR Point Clouds Understanding Network Leveraging Correlated Feature Mining and Geometric-Aware Modelling)
Hardware GTX 1080
Used additional training data 0
Last submission 2021-05-20 17:39:10
Is opensource 1
Number of submissions 7


A_IoU OA [s] IoU 1 IoU 2 IoU 3 IoU 4 IoU 5 IoU 6 IoU 7 IoU 8

Full Confusion Matrix

L1' L2' L3' L4' L5' L6' L7' L8'