code of Team liyiqing_cs on ACMMM24 Multimedia Drone Satellite Matching Challenge In Multiple-environment

code waits to be cleaned

motivated by Generalized UAV Object Detection via Frequency Domain Disentanglement (CVPR 23) , adopt Frequency Distenglement in drone-view localization

• Ours*: train a weather classification network; jointly optimize frequency domain filter and MLPN for domain generaliztion (only drone-view image go through filter, and exists two filter that extract domain invariant spectrum and domain specific spectrum)

• ffm_test.py/ffm_train.py: jointly optimize frequency domain filter and MLPN for domain generaliztion (only drone-view image go through filter, and exists only one filter that extract domain invariant spectrum)

  • ffm_train_iter: use alternating optimiztion in ffm_train

• ffm2.py: jointly optimize frequency domain filter and MLPN for domain generaliztion ( both satllite view image and drone-view image go through filter)

train Image2Image Module to change multi-weather domain image into normal image

• ffm_naive: train frequency filter as Image2Image Module
• repair*: train a toy GAN as Image2Image Module (not working)
• simple_replace.ipynb: as a part to use pix2pix as Image2Image Module (training stage code is not included in this repo)

HOWEVER, we get AP@1 85.08 by using none of ideas above, just by simply using MLPN with augmented University-1652's training set

We included the code of

• LPN

  • ./LPN
  • LPN.ipynb: test LPN on university-1652 and a mixed scenerio
  • LPN-wx.ipynb test LPN on augmented

• MuseNet

  • ./MuseNet
  • MuseNet.ipynb: test MuseNet on mixed scenerio

• MLPN

  • ./MLPN
  • MLPN.ipynb: test MLPN on mixed scenerio
  • MLPN_ensemble.ipynb: train 10 MLPN for each weather case and ensemble them

as baselines and run the related experiments

• ./utils

some common code for all these methods.