- Handle the problem of segmenting objects at multiple scales
- Atrous convolutions in cascade or parallel at multiple atrous rates
- Augment atrous spatial pyramid problem
- Two challenges:
- reduced resolution by consecutive pooling operations: fixed by atrous convolutions
- Remove downsampling operation from last few layers
- And upsampling the corresponding filter kernels
- Equivalent to inserting holes between filter weights
- existence of objects at multiple scales:
- use image pyramid + encoder-decoder + extra modules (DenseCRF) + spatial pyramid pooling
- reduced resolution by consecutive pooling operations: fixed by atrous convolutions
- Atrous spatial pyramid pooling (ASPP) experiment
- Image Pyramid: same model, typically with shared weights, applied to different scales of the image
- Encoder-decoder: encoder where spatial dimension of feature maps is gradually reduced, decoder where object details + dimension are recovered
- Context module: DenseCRF for encoding long-range context
- Spatial pyramid pooling: captures context in several ranges, sometimes based on LSTM
- Atrous convolutions: Experiments with different atrous ranges to capture long-range information
- Max-pooling + striding at consecutive layers reduces the spatial resolution of feature maps in DCNNs
- Atrous convolutions rate allows us to control how densely to compute features in CNNs
- Duplicate several copies of the last ResNet block (block4) and arrange them in cascade
- 3x3 convolutions in these blocks with stride of 2 except the last one
- Four parallel atrous convolutions with different atrous rates applied on top of feature map
