Iterative Siamese Attention Mechanism for 3D Single Object Tracking

The objective of 3D single object tracking is to determine the location of a target in a series of point clouds. However, this task still faces several challenges, such as the sparsity of the point cloud and the timeliness requirement of models. In this letter, we propose using an iterative siamese framework to improve tracking performance. We propose a feature learning backbone that captures correlations between different data samples. Next, we design an iterative siamese attention for feature enhancement of both the template and search to further distinguish the object from the background. Specifically, to perform the information interaction between the two, we employ a cross-attention operation on the template and the search, respectively. To enhance the correlations of the interior of the point cloud, we apply a self-attention operation on them. We repeat this cross-attention and self-attention operation several times to refine the point cloud feature learning. Following the iterative siamese attention module, we input the search feature into the Bird Eye View(BEV) target location network and utilize three prediction heads to predict the target bounding box. Experiments demonstrate that our method consistently outperforms several state-of-the-art approaches in some categories by a significant margin.