<p>Reliable quality control remains one of the most pressing challenges in additive manufacturing (AM), where dimensional inaccuracies can accumulate layer by layer, leading to compromised part performance. While machine learning (ML) has shown promise for predicting defects, its effectiveness is fundamentally limited by the scarcity of standardized, high-resolution datasets that capture layer-wise geometric variations. This paper addresses that gap by introducing a validated, vision-based framework for collecting dimensional accuracy data during fused filament fabrication (FFF). By integrating a calibrated imaging system, a digital reference generator based on G-code, and a robust error quantification pipeline, the framework captures detailed layer-by-layer deviations in real-world units. Validation using physical artifacts confirms the accuracy of the image-based measurements, and the resulting dataset spanning hundreds of layers with corresponding process parameters and thermal readings enables downstream applications in ML-based quality prediction. Predictive models trained on this dataset demonstrated strong performance in forecasting dimensional errors, confirming the effectiveness of the framework. This work lays the foundation for scalable, non-intrusive, and cost-effective data acquisition in AM, facilitating reproducible research and real-time process optimization.</p>