Semantic segmentation with soft cross-entropy loss

Abstract

A system and method for semantic segmentation with a soft cross-entropy loss is provided. The system inputs a first color image to an input layer of a semantic segmentation network for a multi-class classification task. The semantic segmentation network generates, at an auxiliary stride, a first feature map as an output of an auxiliary layer of the semantic segmentation network based on the input first color image. The system extracts the generated first feature map from the auxiliary layer and computes a probability map as a set of soft labels over a set of classes of the multi-class classification task, based on the extracted first feature map. The system further computes an auxiliary cross-entropy loss between the computed probability map and a ground truth probability map for the auxiliary stride and trains the semantic segmentation network for the multi-class classification task based on the computed auxiliary cross-entropy loss.

Description

[0001] CROSS-REFERENCE / INCORPORATION BY REFERENCE TO RELATED APPLICATIONS [0002] This application claims priority to U.S. Provisional Patent Application Serial No. 62 / 758,781, filed November 12, 2018, the entire contents of which are incorporated herein by reference. technical field [0003] Various embodiments of the present disclosure relate to machine learning and computer vision. More specifically, various embodiments of the present disclosure relate to systems and methods for semantic segmentation with soft cross-entropy loss. Background technique [0004] Semantic segmentation is one of the key components of scene understanding, which is the task of assigning semantic labels to individual pixels. Autonomous mobile agents can be used in a wide range of applications, such as self-driving cars, drones and entertainment robots, as well as augmented reality devices and surveillance. These application domains require efficient inference speed and the ability to process h...

AI Technical Summary

Problems solved by technology

However, it remains a challenge to obtain efficient all-in-one models capable of running high-resolution images in resource-constrained computing environments, especially mobile environments operating with limited memory and computing resources.

To achieve a desired level of classification accuracy on high-resolution images, conventional models for semantic segmentation create large parameter sizes and consume significantly large amounts of memory during training time, which is less useful for mobile training environments like autonomous vehicles

Images