monocular-3d

Independent PyTorch implementation for VLMOD multi-object 3D visual grounding from monocular RGB scenes.

Zero-manual-annotation multi-view 3D label-production pipeline for high-occlusion contact sports, plus the monocular lifting recipe built on it — with a layer-wise error budget of what each stage costs

One static phone video of a basketball game to a broadcast that cannot say anything that did not happen. Built end to end against a synthetic fixture; every published number regenerates from a bench script.