Kunfei Li, Youqiang Dong, Bongrae Park, Thomas Koch, and Zhibo Wan
Point cloud retrieval, simultaneous localisation and mapping (SLAM), loop closure, place recognition
Place recognition tasks play a pivotal role in simultaneous localisation and mapping (SLAM). The primary challenge encountered in the place recognition task lies in devising an effective methodology to transform a local descriptor into an informative global descriptor, facilitating accurate retrieval for place recognition. In large-scale point cloud scenes, existing networks often fail to consider the long-range context dependencies between local features and ignore the optimisation problem of local features, leading to additional loss of features. In this paper, we introduce an expand- Laplacian attention network (POE-Net). This network integrates the attention mechanism with the Laplacian process, introducing long-range context dependencies and further optimising to obtain discriminative global descriptors. Our approach exhibits exceptional performance, as evidenced by the experimental outcomes on four benchmark datasets. On the Oxford RobotCar dataset, our average recall at the top 1% reached 88.3%, an 8.0% improvement compared to PointNetVLAD. On the three in-house datasets, our network achieved respective improvements of 10.5% (in the U.S.), 16.8% (in R.A.), and 6.5% (in B.D.) in the average recall at the top 1%, compared to PointNetVLAD.
Important Links:
Go Back
