Investigating Instance Segmentation Performance In Relation to Dataset Composition and Quality

Martell Bell; Rachel V. Vitali

doi:10.1109/ACCESS.2026.3715007

Back

Investigating Instance Segmentation Performance In Relation to Dataset Composition and Quality

Abstract

Open access

Peer reviewed

Investigating Instance Segmentation Performance In Relation to Dataset Composition and Quality

Martell Bell and Rachel V. Vitali

IEEE access, Vol.14, pp.1-1

2026

DOI: 10.1109/ACCESS.2026.3715007

Appears in UI Libraries Support Open Access

Files and links (1)

url

https://doi.org/10.1109/ACCESS.2026.3715007View

Published (Version of record) Open Access

Abstract

High-mix, low-volume (HMLV) manufacturing faces persistent barriers to deploying deep learning-based perception due to the cost and effort of collecting large, labeled image datasets. This study investigates how training dataset composition, specifically the proportion of real images and the use of synthetic and/or augmented supplementation, relates to instance segmentation performance on real images. Using the Sandcast Images for Foundry Automation (SIFA) dataset, a fixed test set of 108 representative real images was extracted and held constant across all trained models. From the remaining real, augmented, and synthetic images, training pools were generated for each supplementation strategy and ratio of real data. For every condition, 3 random seeds were used to split each pool into 756 training images and 216 validation images. YOLOv8 instance segmentation models initialized with MS COCO weights were evaluated on the same real image test set. Objective image quality metrics were computed for each training subset to characterize dataset structure and relate these properties to model fitness. Under the controlled conditions of this study, introducing a small fraction of real data was associated with substantial performance gains, after which improvements diminished. Datasets combining synthetic and augmented images with limited real subsets generally achieved performance comparable to the all real baseline. Overall, the results suggest effective training datasets benefit from structured, complementary variation that remains anchored to the real image domain, highlighting pathways toward more data efficient model development for HMLV applications.

Instance segmentation

image quality

artificial intelligence

neural network

synthetic data

UIOWA OA Agreement

Details

Title: Subtitle: Investigating Instance Segmentation Performance In Relation to Dataset Composition and Quality
Creators: Martell Bell - University of Iowa
Rachel V. Vitali - University of Iowa
Resource Type: Abstract
Publication Details: IEEE access, Vol.14, pp.1-1
DOI: 10.1109/ACCESS.2026.3715007
ISSN: 2169-3536
eISSN: 2169-3536
Publisher: IEEE
Number of pages: 1
Language: English
Date published: 2026
Academic Unit: Roy J. Carver Department of Biomedical Engineering; Mechanical Engineering
Record Identifier: 9985183641202771

Metrics

1 Record Views