The KITTI Vision Benchmark Suite

we propose a weakly supervised 3D detection model that relies solely on readily available \emph{2D} bounding box annotations from images, along with prior information on object size, eliminating the need for costly 3D annotations.
Nevertheless, supervision with 2D bounding-boxes only introduces inherent projection ambiguities in determining the actual 3D object poses.
To this end, for each class, we build a corresponding synthetic class made of basic geometric shapes (ex: 3D boxes) that reproduce the coarse dimensions of their associated real class.
Furthermore, contrary to traditional 3D losses, we notice that 2D supervision losses do not verify range-invariance properties.
For this reason, we introduce a novel range-normalized 2D loss

\alpha=0.2