Comparing Surface Landmine Object Detection Models on a New Drone Flyby Dataset

This is a specialized academic dataset and benchmarking paper addressing a critical humanitarian challenge (landmine detection) that falls outside mainstream computer vision research. The novelty lies in the combination of drone aerial footage + small object detection + landmine-specific domain framing, but uses standard object detection architectures (YOLOv5, Faster R-CNN, etc.). With 0 stars and 2 forks on GitHub, this is an early-stage research contribution with minimal adoption. The project appears to be a reference implementation accompanying a preprint paper (arXiv 2410.19807), not a standalone productized tool. Defensibility is low because: (1) the core models are commodity open-source architectures, (2) no proprietary training methodology is evident, (3) the dataset, while novel and valuable, could be replicated or improved by well-resourced actors (NGOs, militaries, defense contractors). Platform domination risk is medium because cloud providers (AWS, Google, Azure) and defense-oriented AI companies (Palantir, Anduril, Scale AI) could integrate landmine detection into broader geospatial AI platforms. However, displacement horizon is 3+ years because this is a niche humanitarian/defense application without immediate commercial pressure. Market consolidation risk is low—there is no incumbent commercial market for landmine detection ML tools; the space is dominated by traditional geospatial survey and humanitarian demining organizations, not software vendors. The dataset and benchmark are valuable for research reproducibility but lack the ecosystem lock-in or community gravity of production infrastructure projects.