Салата, КирилоАнтощук, Роман2025-09-092025-09-092025https://ekmair.ukma.edu.ua/handle/123456789/36519This research explores how to achieve the minimum sufficient performance for military equipment recognition within a hypothetical, wide-area situationalawareness system. Once demanded deep expertise in neural-network architecture design and complex deployment pipelines can now be realized using two YOLOv8 models built entirely on open-source datasets, frameworks, and free cloud GPUs delivering 69.5% and 72.4% mAP@50 accuracy. We met the problem of merging sources with wildly varying scopes, resolving labeling inconsistencies and duplicate images that inflated confidence scores, addressing severe class imbalance across 14 vehicle categories, and witnessing inter-class confusion among visually similar targets. This work demonstrates that, with modern tools, building a robust military equipment reconnaissance pipeline - from raw images to deployable video inference is both practical and accessible, achieving high accuracy on fine-grained classes while running smoothly on limited GPU resources.en-USrecognitionmilitary equipmentneural-networkcomplex deployment pipelinesterm paperOther