By adding a side network to a face recognition network, output of early convolution blocks may be used to determine relative bounding box values. The relative bounding box values may be used to refine existing boundary box value with an eye on improving the generation, by the face recognition network, of embedding vectors.