Improving Object Detection by Modifying Synthetic Data with Explainable AI

Synthetic data can fill real-world gaps: When real training data is limited or lacks certain examples (like vehicles at specific angles), synthetically generated images can improve model performance by up to 4.6%, demonstrating a practical solution for underrepresented scenarios in computer vision applications.

More realism isn't always better: The research found that both increasing and decreasing realism in synthetic images can improve model performance, challenging the common assumption that synthetic data should simply look as realistic as possible. This suggests designers have more flexibility in how they create training data.

Explainable AI can guide human decision-making: By using XAI saliency maps to show which parts of images the model focuses on, researchers created a "human-in-the-loop" process where people can make informed decisions about which 3D models to modify, reducing the guesswork and expertise required from operators.

Practical improvements are achievable: The framework achieved an additional 1.5% performance improvement beyond the initial synthetic data boost, demonstrating that systematic, AI-guided modification of training data can deliver meaningful real-world gains in detection accuracy.

Addresses a key AI responsibility challenge: By making synthetic data design more efficient and less dependent on specialized expertise, this approach helps democratize the process of improving AI systems, potentially making responsible AI development more accessible to organizations with limited resources.