Rethinking Spatial Encoding Scheme for BEVFormer

In the field of autonomous driving, perception tasks have always been crucial. At present, pure visual perception systems that use visual transformers to fuse multi-view images are receiving increasing attention. However, the spatial perception ability of the existing visual transformer methods is still limited, and it is difficult to match the high requirements of autonomous driving system for perception performance. Given the effectiveness of positional encoding in linking context in the NLP field, we attempt to explore the application of spatial encoding in current visual Transformer models, enabling the models to have the ability to connect spatial contexts. Therefore, this article reviews existing visual position encoding schemes and proposes a spatial encoding design approach. Firstly, the specific tasks of the BEVFormer model were analyzed to identify the spatial attributes that need to be encoded. Then, using group theory, a representation method for spatial attributes was proposed, and two combination structures, serial and parallel, were introduced. The results show that compared with directly using traditional position encoding schemes, the parallel spatial encoding scheme proposed in this paper improves performance by 5.7% and has a faster learning speed in the early stages of training. This study provides a reference for improving the spatial perception ability of visual Transformer models, and in the future, we will continue to explore the effectiveness of spatial encoding for different visual tasks.