Diffusion-Based Synthetic Augmentation for Vehicle Detection

Research Title : Diffusion-Based Synthetic Augmentation with Quality-Aware Reference Selection and mAP-Guided Allocation for Enhancing Vehicle Detection Robustness

Co-Researchers : Dr. Weerachai Wongweeranimit Miss Natnicha Wuttiwaranya Mr.Kerkrith Na. Muengklai Miss Krittiya Phitchakian and Mr. Sarunpong Musikeaw

The study proposes a quality-aware reference selection method based on image resolution, blur, and bounding-box filtering, alongside an mAP-guided allocation strategy that distributes synthetic data according to class-specific performance gaps.

Experiments on a Thai highway surveillance dataset containing 10,105 training images across 14 vehicle classes show that quality-aware reference selection achieves an mAP50 of 84.5% ± 0.6% using 2,768 synthetic images. This outperforms the baseline of 83.2% ± 1.0% and random selection at 83.5%. Under a budget of 2,000 synthetic images, mAP-guided allocation achieves a mean mAP50 of 84.1% ± 0.6%. However, this improvement over the baseline is not statistically significant.