New AI method creates 3D models from single photos using far less training data

Researchers have developed MetaSSP, a new AI method for creating 3D models from single photos. It requires far less manually labeled training data by cleverly using unlabeled images. The system starts with a small 10% labeled dataset, then uses a novel weighting mechanism to learn from unlabeled examples. On a standard benchmark, it reduced shape error by 20.6% and improved accuracy by 24.1%, setting a new performance record.