Method and apparatus with data labeling

An apparatus comprising:

• one or more processors;
• memory storing instructions configured to, when executed by the one or more processors, cause the one or more processors to:
• obtain access to a landmark database comprising landmark points that are geographic coordinates of respective landmarks in a geographic coordinate system;
• after obtaining access to the landmark database, obtain localization information of the moving object, the localization information comprising a geographic location of and/or a direction from the moving object;
• according to the localization information, extract landmark points corresponding to a vicinity of the moving object from the landmark database;
• determine a region of interest (ROI) by determining a horizontal and vertical lengths of an area around the moving object using a bird-eye view image;
• identify one or more visible landmark points located within the determined ROI among the extracted landmark points;
• transform the geographic coordinates of the identified one or more landmark points local coordinates into local coordinates in a local coordinate system of the moving object or a camera thereof;
• generate a ground truth image by projecting the local coordinates of the landmark points into an image domain;
• generate training data by refining the generated ground truth image, the refining comprising searching the ground truth image for target landmark points corresponding to the identified one or more landmark points;
• and
• train a neural network based on the training data.

A method of generating training data performed for a moving object, the method comprising:

• obtaining access to a landmark database comprising landmark points that are geographic coordinates of respective landmarks in a geographic coordinate system;
• after obtaining access to the landmark database, obtain localization information of the moving object, the localization information comprising a geographic location of and/or a direction from the object;
• extracting landmark points corresponding to a vicinity of the moving object from the landmark database according to the localization information;
• generating a ground truth image by projecting the local coordinates of the landmark points into an image domain;
• and
• generate training data by refining the generated ground truth image, the refining comprising searching the ground truth image for target landmark points corresponding to the identified one or more landmark points.