Method, system and apparatus for monocular depth estimation

A system comprising: a server configured to: generate community machine learning models from a global model; distribute each community model to client devices within the corresponding community for training; receive, from each community, weights generated by the training, wherein the model weights indicate structural errors in the model; update the global ML model using the received model weights; and a client device configured to: train a community model based on a training database on the client device; obtain, for an object in an image from the database, a first depth estimation from a sensor and a second depth estimation from the community model; calculate two metrics using the depth estimations; share model weights generated by the training with a server for updating the global model when the first metric indicates structural errors, and other client devices in the community when the second metric indicates there scale errors.