How osupcvlab/pcvlabdrone2021 works
Project type and purpose
• This is a specialized geospatial data-processing pipeline rather than a typical web or mobile full-stack app. It consists of three cooperating Python-based subsystems:
a. “Satellite Map Generation” (offline data collection and tiling)
b. “GIS Mask from OSM” (offline building-mask generation)
c. “UAV Geolocalization” (online/localization pipeline using UAV imagery against prepared map data)
Key top-level components
• Offline Data Preparation
• Satellite Map Generator (create_map.py, main.py, MapAnalysis.ipynb, datasets/)
• GIS Mask Generator (GISgenerator.ipynb)
• Data Storage
• Raw satellite tiles (Satellite Map Generation/datasets/satellite_imgs)
• Coordinate look-up (datasets/coordinates.txt)
• OSM-derived building masks
• Featurebase (UAV Geolocalization/featurebase/)
• Model weights (superpoint_v1.pth, etc.)
• Online Geolocalization Pipeline
• Feature Extraction (superpoint, Feature_extractor.py)
• Feature Matching (SuperGlue, matching.py)
• Quad-Tree Spatial Index for fast geo-lookup
• Pose Estimation (rotation, height modules)
• Output generation (images, videos under output/)
Data-flow and interactions
1. User supplies target area → Satellite Map Generation
2. Output: Tiled satellite imagery + coordinate index
3. GIS Mask Generation against same bounds → building masks
4. Store all artifacts in a shared data repository
• Online UAV Pipeline (right side):
1. UAV camera frames feed into Feature Extractor
2. Queries Quad-Tree index to retrieve corresponding satellite tile & GIS mask
3. Pass features + mask into SuperGlue matcher for putative correspondences
4. Run pose estimation modules (rotation, height) → compute geo-location
5. Log outputs (geoposition, annotated maps, videos)
Architectural patterns and design principles
• Separation of Concerns: clearly split offline data generation vs. real-time localization
• Pipeline stages: each major step is its own Python module/notebook—facilitates testing and swapping algorithms
• Data-driven flow: artifacts flow from storage → retrieval → processing → results
• Use of spatial indexing (Quad-Tree) for O(log n) map lookups
Technologies, frameworks and libraries
• Python 3.8+ scripts and Jupyter notebooks
• OpenCV, NumPy for image processing
• PyTorch for SuperPoint & SuperGlue networks
• OSMnx for OSM queries & building mask generation
• Selenium (chromedriver.exe) for Google Maps scraping
• Geographic indexing via custom Quad-Tree implementation