Semantic Segmentation

Introduction

Semantic segmentation is used to identify the regions and boundaries of objects in images, and is a common problem in the field of computer vision. For semantic segmentation problems, we provide an automated modeling solution that enables users to solve semantic segmentation problems they encounter on their own.

Evaluation Metrics

mIoU

mIoU calculates the ratio of intersection to union between the ground truth and predicted value sets. The specific formula is as follows:

mIoU = \frac{1}{k+1}\sum_{i=0}^{k}{\frac{p_{ii}}{\sum_{j=0}^{k}{p_{ij}}+\sum_{j=0}^{k}{p_{ji}}-p_{ii}}}

$p_{ij}$ represents the number of samples with ground truth $i$ that are predicted as $j$ , and $K+1$ is the number of classes (including the background class). $p_{ii}$ is the number of true positives. $p_{ij}$ and $p_{ji}$ represent false positives and false negatives, respectively.

Parameter Description

Parameter	Description
image_height	(Required) Image height
image_width	(Required) Image width
train_meta	(Required) Training set CSV file path
val_meta	(Required) Validation set CSV file path
test_meta	(Optional) Test set CSV file path
early_stopping_step	(Optional) Early stopping training steps, default is 10
classes	(Required) Names of annotation labels, in array format. Where 0 represents background, and 1~N represent the corresponding label names

Practical Examples

In this section, we will demonstrate the results of several example problems in combination with data formats.

Remote Sensing Land Use Classification

This problem involves identifying land use types in different regions of remote sensing images, with a total of 8 classes. Through the automated semantic segmentation training process, the mIoU result is 0.8521.