arrow detection

Blending synthetic and real data to sharpen arrow detection with ConvNext

Twenty hours of hand-labeling real archery targets. That's the cost of blending 8,000 real images with 40,000 synthetic Blender renders to train a ConvNext DINOv3 detector. The result is worth it: 96.6% F1 on test data and 80.5% full-image correctness with zero missed arrows or false detections. The trick is a two-stage process, registration first, then detection with a clever homography correction for warped paper. This approach to synthetic data echoes the promise of our piece on synthetic time series forecasting accepted at NeurIPS 2026.