EUVIMEDOpdag EUVIMED
Systematische ÜbersichtOpen access · CC BY-NC-SA 4.0

Beyond Segmentation: Bridging the Synthetic-to-Real Gap in Computer Vision A Systematic Review, Taxonomy, and Research Agenda

Roaa Ghanim

Al-Noor Journal of Engineering Management and Computer Science · 2026-08-08 · 71-98

DOI: 10.71229/8chpd013 ↗

At a glance

Systematische Übersicht in Al-Noor Journal of Engineering Management and Computer Science, published online on 2026-08-08. Written by Roaa Ghanim. 101 references. The full text is freely available (open access, CC BY-NC-SA 4.0).

AbstractBibliografiske oplysningerCiteRelated studies

Abstract (original language)

Background

Synthetic imagery produced by simulators, graphics engines and generative models is now widely used in place of real annotation that is costly, privacy-restricted or unobtainable. Models trained on synthetic data nevertheless underperform when deployed on real data, a discrepancy known as the synthetic-to-real gap. The relevant literature is dispersed across domain adaptation, generative data synthesis, domain randomisation, feature alignment, test-time adaptation and evaluation methodology, and no prior synthesis has treated these strands as a single problem space.

Objectives

This review (i) identifies and classifies the sources of the synthetic-to-real gap; (ii) consolidates bridging methods within a common taxonomy in which domain adaptation is one of six families rather than the organising principle of the field; (iii) compares the tasks, datasets, simulators and evaluation protocols used to measure the gap; (iv) appraises the methodological quality and reproducibility of the primary evidence; and (v) proposes an evidence-based research agenda.

Methods

The review was conducted in accordance with the PRISMA 2020 statement [1]. Six bibliographic databases and two preprint repositories were searched to 31 December 2025. Records were screened in duplicate; 68 studies met the eligibility criteria and 34 contributed to the quantitative synthesis. Because no established risk-of-bias instrument is applicable to computer-vision transfer studies, RoB 2 [2] was adapted to this setting. Reported gains are presented descriptively in a forest-style display and are not pooled meta-analytically, because the primary studies report no variance estimates.

Results

On the canonical GTA5 → Cityscapes benchmark, the median improvement over a source-only baseline was +12.9 mIoU (IQR 9.4 to 18.9; k = 10) for convolutional backbones and +28.2 mIoU (IQR 25.4 to 29.2; k = 3) for transformer backbones. Changing the backbone alone, with the adaptation procedure held fixed, raises source-only performance from 34.3 to 45.6 mIoU [3], indicating that a substantial share of the reported "adaptation" gain is architectural. The strongest current method attains 75.9 mIoU against an oracle of 76.4 [4], so the canonical benchmark is effectively saturated while the deployment failures it was intended to predict remain unaddressed. Appearance and feature-level gaps are comparatively well studied, whereas contextual, annotation and sensor gaps are not. Only 13% of studies reported run-to-run variance, and only 25% released the synthetic assets or generation configurations required to regenerate their data.

Conclusions

The synthetic-to-real gap has not been solved; it has been optimised away on a single benchmark. The outstanding challenges are no longer the design of new alignment losses, but the generation of data selected for utility rather than realism, the isolation of architectural from adaptational effects, evaluation in open-world and multi-domain settings, and the release of open simulation pipelines.

Cite

APA

Ghanim, R. (2026). Beyond Segmentation: Bridging the Synthetic-to-Real Gap in Computer Vision A Systematic Review, Taxonomy, and Research Agenda. Al-Noor Journal of Engineering Management and Computer Science, 71-98. https://doi.org/10.71229/8chpd013

Vancouver

Ghanim R. Beyond Segmentation: Bridging the Synthetic-to-Real Gap in Computer Vision A Systematic Review, Taxonomy, and Research Agenda. Al-Noor Journal of Engineering Management and Computer Science. 2026:71-98. doi:10.71229/8chpd013

BibTeX
@article{Ghanim2026,
  author = {Ghanim, Roaa},
  title = {Beyond Segmentation: Bridging the Synthetic-to-Real Gap in Computer Vision A Systematic Review, Taxonomy, and Research Agenda},
  journal = {Al-Noor Journal of Engineering Management and Computer Science},
  year = {2026},
  pages = {71--98},
  doi = {10.71229/8chpd013}
}
Start for free →