Depth Perception Using Various Vision Transformer

Proper depth perception is one of the key requirements of three-dimensional understanding of scenes in the context of self-driving. The discussed manuscript defines a re-architecturing of VoxelNet with a dual attention paradigm (inspired by Vision Transformers (ViT)) added to capture long-range relationships and context-sensitive features. The combined use of channel-wise and location attention modules in encoding voxel features produces improvements in the effectiveness of object characterization and location of objects. Empirical analyses performed on the KITTI 3D dataset demonstrates that there can be observed better results in depth-perception accuracy and mean average-precision in the pedestrian, cyclist and vehicular categories.