ACMMM23-Solution-MBEG

Paper Link: https://doi.acm.org/?doi=3607834.3616562

Modern Backbone for Efficient Geo-localization

Code Overview:

• config: settings.yaml

• model definitions: model_.py

• train: train.py

• distillation knowledge: distill_train.py

• test: U1652_test_and_evaluate.py

• prepare dataset: Preprocessing.py

• multiply queries: multi.py

• draw heat map: draw_cam_ViT.py

• predict University160k : predict.py / predict.ipynb

• export answer.txt : export.ipynb

Model weights

Baidu Cloud Disk Link: https://pan.baidu.com/s/1k1z90EyLaL85PqeSxxlWOw?pwd=1652 提取码: 1652

• MBEG-L1 for University-1652: MBEG-L1-1652.pth

  • Drone -> Satellite: Recall@1: 92.50 AP: 93.75
  • Satellite -> Drone: Recall@1: 94.15 AP: 91.57

• MBEG-L2 for University-1652: MBEG-L2-1652.pth

  • Drone -> Satellite: Recall@1: 89.78 AP: 91.53
  • Satellite -> Drone: Recall@1: 92.01 AP: 88.81

• MBEG-B1 for University-1652: MBEG-B1-1652.pth

  • Drone -> Satellite: Recall@1: 87.48 AP: 89.36
  • Satellite -> Drone: Recall@1: 90.73 AP: 86.29

• MBEG-B2 for University-1652: MBEG-B2-1652.pth

  • Drone -> Satellite: Recall@1: 88.16 AP: 89.97
  • Satellite -> Drone: Recall@1: 93.01 AP: 87.64

• MoblieViT-KD for University-1652: MobileViT-Student-1652.pth

  • Drone -> Satellite: Recall@1: 80.57 AP: 83.44
  • Satellite -> Drone: Recall@1: 88.44 AP: 80.45

Related Works

University-1652 Benchmark

https://github.com/layumi/University1652-Baseline

ACM MM23 Workshop: UAVs in Multimedia: Capturing the World from a New Perspective

https://www.zdzheng.xyz/ACMMM2023Workshop

EVA Series backbone

https://github.com/baaivision/EVA https://huggingface.co/Yuxin-CV/EVA-02

MobileViT

https://github.com/apple/ml-cvnets [https://huggingface.co/timm/mobilevitv2_200.cvnets_in22k_ft_in1k_384] (https://huggingface.co/timm/mobilevitv2_200.cvnets_in22k_ft_in1k_384)