The KITTI Vision Benchmark Suite

3D point cloud annotation is laborious, spurring
the development of 3D automatic annotation.
However, to preserve the high-quality 3D
annotation, current methods usually introduce
complicated annotation pipelines, e.g., staged
training for 3D foreground/background
segmentation, cylindrical object proposals, or
point completion. We propose a simple yet
effective end-to-end Context-Aware Transformer
(CAT) as an automated 3D-box labeler, only
requiring a small set of labeled 3D point clouds
and weak 2D boxes. Compared to existing annotation
methods, CAT requires minimal modifications to the
vanilla Transformer block. We adopt a Transformer
encoder consisting of local and global parts and a
standard Transformer decoder. Specifically, the
local encoder explicitly models long-range
contexts among points at an object level, and the
global encoder captures contextual interactions
across objects for point feature learning.

depth=[6,3,3]
embed_dim=512
num_points=1024