The KITTI Vision Benchmark Suite

MonoLDC detects 3D objects from a single
calibrated RGB image. It augments the detector's
feature pyramid with depth-aware visual features.
In later decoder layers, the preceding layer's box
predictions guide object-specific sampling from a
shared multiscale depth memory. The current box
geometry and boundary uncertainty condition a
learned residual that refines a projection-based
depth estimate.

SSLD-pretrained ResNet-50, input size = 1280 x 384,
top-k = 50, classification threshold = 0.20.
Runtime measured at batch size 8 on one RTX 4090.