Research motivation :
For semantic segmentation we use zero padding to prevent feature map will diminish fast,but the and also lose the border information.
The insufficient Zero padding : for the learned filter(include learn from border will share in all spatial locations)
So ,they want to use context-aware (CA) to padding:
To know what the first layer receive,only taking local region to directly predict the displacement to extrapolate the image. (the image extrapolation)
The main contributions:
- new padding method according to image extrapolation.
- Compare with sota method in SS task.
Method
Network:
Algorithm:
Assume the as the input image
Expermients:
Datasets | Cityscape | DeepGlobe satellites |
total | 5000 | 1146 |
train_set | 2975 | 803 |
test_set | 1525 | 172 |
vaildation set | 500 | 171 |
numer_calss | 19 | 7 |
Metric: mlou
Segmeantation network :PSPNet
Conclusion :
higher mlou than zero padding.
The point:
pyramid pooling : Multi-scale features are extracted from fixed size feature vectors.
Know the difference of SPP-Net and R-CNN
atrous spatial pooling: first be put forward in Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs in the condition of keeping feature map sable to increase receptive-field
Image regeneration :
- GANs based model
- autoregressive models
The padding way:
- zero padding
- Reflection padding
- circular padding ---- >re-weighting based scheme called prartial convolution
- mean distribution padding
- symmetric padding
The inspiration:
to care more border information can lead to better segmentation result.the moudle maybe can be used !!!!