Method and apparatus for training an object detector with integration of multiple detection head

Abstract

A computer-implemented method for training a neural network (60) wherein the neural network (60) is designed to detect an object in an image (x) and the training comprises the steps of: determining a feature representation (f) of a training image (xi) by means of a trunk (B) of the neural network (60); determining a head output (C1, C2, C3, P1, P2, P3, S1, S2, S3) on the basis of the feature representation (f) and by means of a plurality of (MC, MP, MS) heads (HC1, HC2, HC3, HP1, HP2, HP3, HS1, HS2, HS3) of the neural network (60), the head outputs (C1, C2, C3, P1, P2, P3, S1, S2, S3) characterizing the presence or absence of an object, or characterizing a range of the object, or characterizing a position of the object, or characterizing the range and the position, respectively, for a region of the training image (xi); determining a plurality of first loss values, the first loss values respectively corresponding to the head outputs (C1, C2, C3, P1, P2, P3, S1, S2, S3) and characterizing a deviation of the head outputs (C1, C2, C3, P1, P2, P3, S1, S2, S3) from a desired head output; adjusting at least one parameter (theta) of the trunk (B) and / or at least one parameter of the head (HC1, HC2, HC3, HP1, HP2, HP3, HS1, HS2, HS3) on the basis of a second loss value characterizing the sum of the first loss values.

Images