README.md

Drone Landing Site Safety

Analyze aerial RGB imagery to detect safe drone landing sites. Combines monocular depth estimation, promptable hazard segmentation, and geometric heuristics to flag flat, obstacle-free areas, with overlays and metrics that show why a spot is safe. A Gradio UI and curated gallery are included for quick testing and browsing of precomputed outputs.

What’s inside

• Main app (landing_app.py) — runs full inference with adjustable thresholds, overlays, and camera assumptions; requires >8GB VRAM (assuming default 1024 px processing resolution); runtime is ~1000ms per image.
• Curated gallery (demo/demo_app.py) — precomputed PNG/JPG/JSON artifacts for fast, zero-GPU browsing.

Prereqs

• Python 3.10+ and a CUDA GPU for the main app (CPU works but is slow).
• Sample images: drop your RGBs under data/Image/. 5 VISLOC images are bundled as examples.
• Install deps: pip install -r requirements.txt.

References

• UAV-VisLoc dataset: Xu et al., 2024 (https://arxiv.org/abs/2405.11936)
• Depth Anything 3: Lin et al., 2025 (https://arxiv.org/abs/2511.10647)
• SAM 3: “SAM 3: Segment Anything with Concepts” (https://ai.meta.com/research/publications/sam-3-segment-anything-with-concepts/)