Aligning Unimodal Encoders with 100-500 Paired Samples via Optimal Transport

This method aligns independently pretrained unimodal encoders using only 100 to 500 paired anchor samples and abundant unpaired data. It uses an entropic optimal transport problem to match point clouds while applying marginal constraints to prevent hubness, finally distilling the mapping into lightweight encoders.