(6th) Solution for 2024 IEEE GRSS Data Fusion Contest Track 1

train data structure

├── data
│   ├── dev
│   │   ├── p1
│   │   └── p2
│   └── Track1
│       ├── train
│       │   ├── images
│       │   └── labels
│       └── val
│           └── images
|
|
└── waterflow
      |....

Run the code

pip install -r requirements.txt

# infer
python run.py

Method

Model = Unetpp + ResNeSt 269e

Loss = Diceloss + BCEloss -> rate=3:1

Pre-process = Normalize(img_mean, img_std)

Output = sigmoid + FindBest(threshold)

Trick = TTA

Result

lr=3e-5, wd=4e-3, warm-up=0.2, ep=400

model\threshold	0.5	0.3	0.1
upp 269e	0.9210
upp 269e + TTA	0.9240	0.92560	0.9260

TODO

• K-Fold
• Visualize
• pre-process upgrade
• ...