Show HN: Trained YOLOX from scratch to avoid Ultralytics (iOS aircraft detect) (austinsnerdythings.com)

Austin-based developer has designed an AR app called SkySpottr to display real-time aircraft information on iPhones. To power his app, he opted to train the YOLOX model from scratch after discovering the licensing restrictions of Ultralytics' YOLOv8. While Ultralytics’ model is well-regarded for its speed and accuracy, its AGPL-3.0 license would require open-sourcing his app or acquiring a costly commercial license. YOLOX, licensed under MIT, allows greater freedom, albeit necessitating the training of models independently without pre-trained weights. The developer successfully trained four variations of YOLOX models specifically for aircraft detection, overcoming early training hurdles by using the COCO2017 dataset. Although initial models struggled with detecting small aircraft at altitude, adjustments like increasing input resolution and optimizing data augmentation led to improved accuracy—showing a detection rate of up to 92.9% for larger models. The lightweight nature of YOLOX configurations such as nanoish enables feasible deployment on mobile devices, balancing performance and accuracy in aircraft detection, crucial for enhancing the user experience in augmented reality applications.