The KITTI Vision Benchmark Suite

we propose a novel Geometric Consistency
Constraint Loss (GCCL) for monocular 3D object
detection. GCCL integrates two key components.
First, a 3D Corner Point Alignment Loss enforces
geometric coherence by calculating the coordinates
of the predicted 3D bounding box's eight corners
in the camera coordinate system and aligning them
with their ground truth counterparts. Second, a
3D-2D Projection Alignment Loss reduces prediction
offset by ensuring the projected 3D box aligns
tightly within its corresponding 2D detection
bounding box on the image plane.

\alpha=0.2

None