Code for Multi-Temporal Scene Classification and Scene Change Detection with Correlation based Fusion. To acquire dataset, please contact [email protected].
We proposed a unified network called CorrFusionNet for scene change detection. The proposed CorrFusionNet firstly extracts the features of the bi-temporal inputs with deep convolutional networks. Then the extracted features will be projected into a lower dimension space to computed the instance level canonical correlation. The cross-temporal fusion will be performed based on the computed correlation in the CorrFusion module. The final scene classification and scene change results are obtained with softmax activation layers. In the objective function, we introduced a new formulation for calculating the temporal correlation. The visual results and quantitative assessments both demonstrated that our proposed CorrFusionNet could outperform other scene change detection methods and some state-of-the-art methods for image classification.
- The proposed CorrFusion module:
- The proposed CorrFusionNet:
scipy==1.1.0
matplotlib==3.0.3
h5py==2.8.0
numpy==1.16.3
tensorflow_gpu==1.8.0
Pillow==6.2.1
scikit_learn==0.21.3
- Overview of our Wuhan dataset
The images are stored in npz format.
├─trn
│ 0-5000.npz
│ 10000-15000.npz
│ 15000-16488.npz
│ 5000-10000.npz
│
├─tst
│ 0-4712.npz
│
└─val
0-2355.npz
pip install -r requirements.txt
python train_cnn.py [-h] [-g GPU] [-b BATCH_SIZE] [-e EPOCHES]
[-n NUM_CLASSES] [-tb USE_TFBOARD] [-sm SAVE_MODEL]
[-log SAVE_LOG] [-trn TRN_DIR] [-tst TST_DIR]
[-val VAL_DIR] [-lpath LOG_PATH] [-mpath MODEL_PATH]
[-tbpath TB_PATH] [-rpath RESULT_PATH]
(see parser.py)
-
Download a trained model here.
-
Evaluation
python evaluate_model.py [-h] [-g GPU] [-m MODEL_DIR] [-tst TST_DIR]
[-val VAL_DIR]
optional arguments:
-h, --help show this help message and exit
-g GPU, --gpu GPU gpu device ID
-m MODEL_DIR, --model_dir MODEL_DIR
model directory
-tst TST_DIR, --tst_dir TST_DIR
testing file dir
-val VAL_DIR, --val_dir VAL_DIR
validation file dir
- The results of quantitative assessments:
- Predictions on our dataset:
For any questions, you're welcomed to contact Lixiang Ru.