Drone-view building identification by cross-view Triplet deep neural network and relative spatial estimation
Demo video: https://www.youtube.com/watch?v=sdq31ep2zYk
-
Environment: tensorflow 0.10.0 numpy 1.11.0 python 2.7
-
Run code:
train triplet: python train_cross.py --model_dir model/ig_21.61 (fine-tuned from --model_dir model_dir09el/ig_21.61)
extract deep features: python extract_triplet_cross.py --model_dir model/weatherB_best (extract fc features by model/weatherB_best )
retrieval:
python retrieval.py --f frame/all
(retrieval in frame/all)
similarity.py: compute similiarty
- Dataset https://jcwchen.github.io/DVBI/
(1) Drone-BR 80 drone-view images: frame/all/ Sensor information per image: location.txt Building information frome Google Place API per image: poi/[name].txt
Building information from Web: a. ground-level image: search/ b. street-view image: streetview_clean/ c. aerial image: aerial_clean/
(2) Drone-BD bounding box per image: Drone-BD/[name].txt
(3) IG-City8 same buildings are in same directory
Download model: https://www.dropbox.com/sh/t8q70qnz9njtytq/AAAE2Hbw159p7Dj86QL2Z9hia