Input Ranking

Pipeline for ranking street-level Mapillary images by building match, with optional scene recognition and segmentation.

Project structure

input_ranking/
├── data/                    # JSON data (input & output)
│   ├── m_out.json          # Raw OSM + Mapillary fetch
│   ├── accepted.json       # Filtered images
│   ├── rejected.json
│   ├── building_rankings.json
│   ├── sequence_rankings.json
│   └── accepted_with_analysis.json
├── output/
│   ├── thumbnails/         # Downloaded images (gitignored)
│   ├── scene_scores/       # ResNet Places365 results
│   ├── segmentation/       # Mask2Former results
│   └── galleries/          # HTML galleries
├── models/
│   └── places365/          # ResNet weights (download separately)
├── *.py                    # Scripts
├── .env                    # MAPILLARY_ACCESS_TOKEN (gitignored)
└── requirements.txt

Pipeline

1. Fetch data: python get_mapillary.py → data/m_out.json
2. Filter: python filter_metadata.py → data/accepted.json, data/rejected.json
3. Rank: python build_rankings.py → data/building_rankings.json, galleries
4. Analyze: python run_analysis.py → data/accepted_with_analysis.json
5. Gallery: python generate_analysis_gallery.py → output/galleries/
6. Visualize: python visualize.py → output/galleries/building_gallery.html

Setup

• Add MAPILLARY_ACCESS_TOKEN to .env
• Place resnet18_places365.pth.tar in models/places365/ for scene recognition