A Comparison of Training Strategies Using Real and Synthetic Data for Cattle Detection

Collecting and annotating representative aerial imagery remains a constraint on automated cattle monitoring. We evaluate simulation-generated images as a complementary training resource for cattle detection from unmanned aerial vehicles (UAVs). A pipeline developed in Unreal Engine 5.3.2 combines randomized agricultural scenes, herd movement, asynchronous animal animations, camera variation, and automatic bounding-box annotation. Quality filtering retained 2084 of 6986 rendered images. We compared real-only training (A), sequential synthetic-to-real training (B), and real:synthetic mixtures of 50:50 (C50) and 75:25 (C75) across five YOLOv8n training seeds, with 2000 optimizer updates per strategy. On the 93-image test split, mean mAP50–95 ± sample standard deviation was 0.5153 ± 0.0063 for A, 0.5083 ± 0.0092 for B, 0.4898 ± 0.0066 for C50, and 0.5027 ± 0.0041 for C75. C75 improved on C50, while real-only training retained the highest mean mAP50–95 under the evaluated protocol. Validation-selected thresholds and counting errors characterized the balance between false detections and misses. A separate three-seed comparison of static and animated synthetic scenes showed fewer misses with animation and movement, alongside more false detections on average, while the mAP50–95 differences varied across seeds. The pipeline provides an annotated-data resource and controlled comparisons of its use in training. Our findings concern fixed synthetic subsets and the existing CowDetect partition. Independent acquisitions and separate animation and movement conditions remain to be evaluated.