Context-Aware Transformer for 3D Point Cloud Automatic Annotation

3D automatic annotation has received increased attention since manually annotating point clouds is laborious. However, existing methods are usually complicated, e.g., pipelined training for foreground/background segmentation, cylindrical object proposals, and completion. Furthermore, they often overlook the inter-object feature correlation that particularly informative to hard samples annotation. To this end, we propose a simple yet effective end-to-end Context-Aware Transformer (CAT) as an automated 3D-box labeler generate precise box annotations from 2D boxes, trained with small number of human annotations. We adopt general encoder-decoder architecture, where CAT encoder consists intra-object (local) (global), performing self-attention along sequence batch dimensions, respectively. The former models interactions among points latter extracts relations different objects, thus boosting scene-level understanding. Via local global encoders, can high-quality streamlined workflow, allowing it outperform state-of-the-arts by up 1.79% AP on task KITTI test set.