Rare-Class Label Expansion for Image Segmentation Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Image segmentation for autonomous vehicles faces challenges with rare classes like caution tape and firehose, as direct down-sampling results in insufficient pixels for training, leading to poor model performance.
Innovation Solution
Apply Gaussian smoothing or dilation functions to expand labels for rare classes, weighting scores by inverse frequency, and down-sample to required resolution for training deep learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If direct down-sampling is applied to image segmentation labels, then processing efficiency is improved, but the number of pixels for rare classes becomes insufficient for training
Solution Approach 1:
The patent applies Gaussian smoothing or dilation functions to expand labels for rare classes before down-sampling. This preliminary expansion ensures that even after down-sampling, sufficient pixels remain for effective training of rare classes like caution tape and firehose
2Quantity of substance
If label expansion is applied to rare classes, then the number of pixels for training is increased, but computational complexity increases
Solution Approach 1:
The patent selectively applies label expansion techniques (Gaussian smoothing or dilation) only to rare classes identified by inverse frequency weighting, rather than processing all classes uniformly. This localized approach increases pixel quantity for rare classes while minimizing unnecessary computational overhead for common classes
3Reliability
If Gaussian smoothing or dilation is applied to expand labels, then model performance on rare classes improves, but boundary precision may be reduced
Solution Approach 1:
The patent applies Gaussian smoothing or dilation functions with controlled parameters to expand labels for rare classes. The expansion is partial and controlled, applying only enough to ensure sufficient training pixels while minimizing excessive boundary expansion that would degrade precision
Data Source
AI summary
An apparatus for facilitating image segmentation on a dataset comprising a plurality of classes is described and includes a module executable by a processor to preprocess the dataset by applying a filter to labels of the plurality of classes to expand the labels of the selected class; down-sample the preprocessed dataset to a desired resolution for training an image segmentation model; and output the down-sampled preprocessed dataset to the image segmentation module, wherein the labels comprise one-hot labels.