Dynamic delta transformations for segmentation

An apparatus for generating one or more segmentation masks, the apparatus comprising:

• at least one memory;
• and
• at least one processor coupled to the at least one memory and configured to:
• generate, using a machine learning system, a delta image based on a difference between first pixels associated with a current image and second pixels associated with a prior image at a plurality of spatial locations;
• combine the delta image with features in a first feature map generated by a machine learning model of the machine learning system representing the prior image to generate a combined feature representation of the prior image;
• process, using a transform operation, the combined feature representation to generate a transformed feature representation of the prior image, wherein the transformed feature representation of the prior image aligns a first feature of an object of the prior image with a second feature of the object of the current image;
• combine the transformed feature representation of the prior image with features in a second feature map generated by the machine learning model from the current image to generate a combined feature representation of the current image;
• and
• generate, using the machine learning system, based on the combined feature representation of the current image, a segmentation mask for the current image, wherein the segmentation mask identifies pixels of the current image that correspond to one or more objects present in the current image.

A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to:

• generate, via a machine learning system, a delta image based on a difference between first pixels associated with a current image and second pixels associated with a prior image at a plurality of spatial locations;
• combine the delta image with features in a first feature map generated by a machine learning model of the machine learning system from the prior image to generate a combined feature representation of the prior image;
• process, using a transform operation, the combined feature representation to generate a transformed feature representation of the prior image, wherein the transformed feature representation of the prior image aligns a first feature of an object of the prior image with a second feature of the object of the current image;
• combine the transformed feature representation of the prior image with features in a second feature map generated by the machine learning model from the current image to generate a combined feature representation of the current image;
• and
• generate, using the machine learning model and based on the combined feature representation of the current image, a segmentation mask for the current image, wherein the segmentation mask identifies pixels of the current image that correspond to one or more objects presented in the current image.