查看文章

In contrast to traditional transformer blocks using a set of pre-defined parameters as positional embeddings, we propose the input-aware positional embedding (IPE) which is dynamically generated according to the input feature. We implement this idea by designing the IPE transformer, which enjoys stronger generalization powers across arbitrary input sizes. To verify its effectiveness, we integrate the newly-designed transformer into NLSPN and GuideNet, two remarkable depth completion networks. The experimental result on a large scale outdoor depth completion dataset shows that the proposed transformer can effectively model long-range dependency with a manageable memory overhead.