Abstract: In one implementation, a method of training a type-agnostic object segmentation system is performed in a type-agnostic object segmentation system including one or more processors, and a non-transitory memory. The method includes obtaining a data set; generating a respective embedding vector for each of a plurality of pixels of the image frame; mapping a plurality of embedding vector representations into a dataspace defined by a dimensionality of the plurality of embedding vector representations; comparing the at least one object instance representation candidate against the respective segmentation mask; and adjusting the type-agnostic object segmentation system in order to satisfy an error threshold across the plurality of image data frames according to a determination that the at least one object instance representation candidate and the respective segmentation mask differ by a threshold value.