The Impact of Data Quality Issues in Computer Vision Data Fusion Models скачать в хорошем качестве

The Impact of Data Quality Issues in Computer Vision Data Fusion Models

Speaker: Daniel Sobien This presentation was part of the SEI's virtual workshop: AI Data Quality: Advancing AI Engineering for Reliable Data Pipelines on October 23, 2025. This talk discusses the impacts of limited size image datasets on Computer Vision (CV) data fusion models. Data fusion is the process of combining disparate data sources or multiple model predictions into a single outcome. I will discuss this method with two contexts for this talk. The first cover simulated Synthetic Aperture Radar (SAR), because it is expensive to collect, but simulating data leads to a Simulated-to-Real gap. Data fusion can help but is susceptible to poor-quality sources, so here I show how Test and Evaluation (T&E) is key to assessing and handling multiple sources for fusion. The second covers limited size satellite image datasets, but small datasets make CV models difficult to train. Again, I use data fusion to improve performance, but here I show how Discount Factors, based on model T&E, can mitigate impacts of fusing results from underperforming models. I focus on how T&E with baseline CV models is critical for understanding data quality issues – either a poor data source or underfit model from limited data. Insights from the T&E process allow us to make improvements on model performance by adjusting the model fusion process.