Awesome
Description
This branch implements a building extraction on remote sensing images, combining the adversarial networks with a FC-DenseNet model.
Network structure
Overview of our segmentation architecture with the adversarial network. Left: The segmentation network takes an aerial image as input and produces a pixel-wise classification label map. Right: A label map, chosen from segmentation output or ground truth, is multiplied with their corresponding input aerial image to produce a masked image, and the adversarial network takes this masked image map as input and adopts an auto-encoder network to reconstruct it.
Preparation
Train
nohup /home/mmvc/anaconda2/envs/Xiang_Li3/bin/python inria3.py --exp_id 1 --model 'FC-DenseNet158' > log1.log
Test
step 1: to get the prediction results python run_prediction.py
step 1: evaluation python eval_aerial.py prediction_#108
Results
Test accuracy of different models on the Massachuttes dataset.
| Model | Breakeven ($\rho$ = 3) | Breakeven ($\rho$ = 0) | Time (s) |
|---|---|---|---|
| Mnih-CNN~\cite{mnih2013machine} | 92.71 | 76.61 | 8.7 |
| Mnih-CNN+CRF~\cite{mnih2013machine} | 92.82 | 76.38 | 26.6 |
| Saito-multi-MA~\cite{saito2016multiple} | 95.03 | 78.73 | 67.7 |
| Saito-multi-MACIS~\cite{saito2016multiple} | 95.09 | 78.72 | 67.8 |
| HF-FCN~\cite{zuo2016hf} | 96.43 | 84.24 | 1.07 |
| Ours (56 layers) | 96.40 | 83.17 | 1.01 |
| Ours (158 layers) | 96.78 | 84.79 | 4.38 |
Validation accuracy of different network depths on Inria Aerial Image Labeling dataset.
| FC-DenseNet (56 layers) | 74.64 | 96.01 |
|---|---|---|
| Ours (56 layers) | 74.75 | 96.01 |
| FC-DenseNet (103 layers) | 75.58 | 96.19 |
| Ours (103 layers) | 76.31 | 96.32 |
| FC-DenseNet (158 layers) | 77.11 | 96.45 |
| Ours (158 layers) | 78.73 | 96.71 |
