Semantic segmentation ground truth correction with spatial transformer networks

An apparatus comprising at least one processor and at least one memory storing computer program code, the at least one memory and the computer program code are configured to, with the processor, cause the apparatus to at least:

• access label data stored in computer-readable memory and corresponding to a geographic area;
• access a plurality of training images corresponding to at least a portion of the geographic area;
• provide the label data and the plurality of training images to a training model that comprises of at least a predictor model and an alignment model, wherein (a) the predictor model is configured to receive at least one image of the plurality of training images and output at least one prediction based at least in part on the at least one image, and (b) the alignment model is configured to generate a transformed prediction based at least in part on the at least one prediction and the label data;
• and
• cause a loss engine to iteratively:
• receive the label data and the transformed prediction from the alignment model,
• evaluate a loss function based at least in part on the label data and the transformed prediction, and
• cause weights of the predictor model and the alignment model to be updated based on the evaluated loss function to cause the predictor model and the alignment model to be trained.

A computer program product comprising at least one non-transitory computer-readable storage medium having computer-executable program code instructions stored therein, the computer-executable program code instructions comprise program code instructions configured to, when executed by a processor of an apparatus, cause the apparatus to:

• access label data stored in computer-readable memory and corresponding to a geographic area;
• access a plurality of training images corresponding to at least a portion of the geographic area;
• provide the label data and the plurality of training images to a training model that comprises of at least a predictor model and an alignment model, wherein (a) the predictor model is configured to receive at least one image of the plurality of training images and output at least one prediction based at least in part on the at least one image, and (b) the alignment model is configured to generate a transformed prediction based at least in part on the at least one prediction and the label data;
• and
• cause a loss engine to iteratively:
• receive the label data and the transformed prediction from the alignment model,
• evaluate a loss function based at least in part on the label data and the transformed prediction, and
• cause weights of the predictor model and the alignment model to be updated based on the evaluated loss function to cause the predictor model and the alignment model to be trained.