Incorporation Of Rejection Criterion - A Novel Technique For Evaluating Semantic Segmentation Systems

Semantic segmentation 'SS' evaluation metrics are great tools to assess systems' performance in terms of pixels' accuracy and the alignment of segments. SS systems represent the core element of many human-system interaction applications. Thus, improving the comprehensiveness of S...

Full description

Saved in:
Bibliographic Details
Main Authors: Mohamed, Elhassan, Sirlantzis, Konstantinos, Howells, Gareth
Format: Conference Proceeding
Language:English
Subjects:
Convolutional Neural Network
Convolutional neural networks
Evaluation metrics
Image segmentation
Measurement
Pixels classification
Semantic segmentation
Semantics
System performance
Task analysis
Thresholding
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Semantic segmentation 'SS' evaluation metrics are great tools to assess systems' performance in terms of pixels' accuracy and the alignment of segments. SS systems represent the core element of many human-system interaction applications. Thus, improving the comprehensiveness of SS evaluation metrics can accurately reflect the performance of human-system inter-action. Standard metrics ignore pixels' confidence scores which can carry useful information. Pixels' scores represent the level of confidence of the system for assigning class labels to image pixels. However, it has not been utilised by any evaluating metric for semantic segmentation systems. We propose to incorporate pixels' confidence scores with existing metrics to gain better insights into systems' behaviours. Results show the usefulness of the introduced approach to utilise the pixels' scores in the evaluation process. Besides, using pixels' scores thresholding can help to enhance the system performance on a specific task or objects of a particular size.
ISSN:2158-2254
DOI:10.1109/HSI52170.2021.9538787