Prelabeling for semantic segmentation tasks

A non-transitory computer readable medium storing instructions that, when executed by a processor, cause the processor to perform the steps of:

• applying a first set of scale factors to an image to produce a set of downscaled versions of the image;
• applying a second set of scale factors to the image to produce a set of upscaled versions of the image;
• performing one or more operations that apply a machine learning model to the image, the set of downscaled versions of the image, and the set of upscaled versions of the image to produce a first set of predicted labels for a first set of pixels associated with the image, a second set of predicted labels for a second set of pixels associated with the set of downscaled versions of the image, and a third set of predicted labels for a third set of pixels associated with the set of upscaled versions of the image;
• generating a semantic segmentation that includes a fourth set of predicted labels for regions of pixels in the image based on an aggregation of the first set of predicted labels, the second set of predicted labels, and the third set of predicted labels;
• and
• outputting, in a user interface, the semantic segmentation in association with the pixels in the image.

A system, comprising:

• a memory that stores instructions, and
• a processor that is coupled to the memory and, when executing the instructions, is configured to:
• apply a first set of scale factors to an image to produce a set of downscaled versions of the image;
• apply a second set of scale factors to the image to produce a set of upscaled versions of the image;
• perform one or more operations that apply a machine learning model to the image, the set of downscaled versions of the image, and the set of upscaled versions of the image to produce a first set of predicted labels for a first set of pixels associated with the image, a second set of predicted labels for a second set of pixels associated with the set of downscaled versions of the image, and a third set of predicted labels for a third set of pixels associated with the set of upscaled versions of the image;
• generate a semantic segmentation that includes a fourth set of predicted labels for regions of pixels in the image based on an aggregation of the first set of predicted labels, the second set of predicted labels, and the third set of predicted labels;
• output, in a user interface, the semantic segmentation for use in assisting a user in specifying updates to the fourth set of predicted labels for the pixels in the image;
• receive, via the user interface, user input indicating locations of one or more labels in the image;
• and
• perform one or more additional operations that apply a second machine learning model to the locations of the one or more labels to update the semantic segmentation.