Reproducibility Package for "A Ceiling-Aware Decision Framework for Assessing the Marginal Utility of Escalating Image Data Augmentation
Year: 2026
Type: software
Abstract: This reproducibility package accompanies the manuscript “A Ceiling-Aware Decision Framework for Assessing the Marginal Utility of Escalating Image Data Augmentation.” The study evaluates whether increasingly sophisticated image-data augmentation strategies provide reproducible marginal utility when a strong baseline already operates near the performance ceiling. Four progressively more complex conditions are examined under a frozen paired multi-seed design: C0, a real-data baseline; C1, conventional image augmentation; C2, StyleGAN2-ADA-based synthetic augmentation; and C3, diffusion-based synthetic augmentation. The archive contains non-image reproducibility materials supporting the reported validation-level analyses, including statistical analysis scripts, figure and table generation scripts, classifier training scripts, synthetic-data quality-control and remediation scripts, primary statistical summaries, primary C3 seed-level metrics, confirmatory C0–C3 seed-level metrics, prespecified primary and confirmatory seed definitions, and SHA-256 provenance bindings. The primary cohort uses seeds 1, 11, 29, 47, 83, 131, 173, 227, 281, and 353. A separately frozen, non-overlapping confirmatory cohort uses seeds 401, 457, 503, 563, 619, 677, 739, 809, 877, and 947. The underlying retail imagery is not redistributed because redistribution rights are not asserted by the authors. Large per-image quality-control inventories are retained privately and are cryptographically bound through SHA-256 digests included in the archive. The documented held-out partition is not provided as an evaluable test artifact because its authoritative frozen manifest provenance could not be re-established at submission finalization; no replacement test partition was constructed post hoc. This release is intended to support transparent verification of the reported analysis workflow, synthetic-data quality-control procedures, paired multi-seed comparisons, and confirmatory seed-based reproducibility assessment. It does not constitute an external-dataset replication or an independent held-out test release.
Institution: Rangsit University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224930670