Deep CNN for other tasks, however there are some fatal flaws. More famous is up-sampling and pooling layer design. In this speech in Hinton also been mentioned.
The main problems are:
- Up-sampling / pooling layer (e.g. bilinear interpolation) is deterministic. (a.k.a. not learnable)
- Internal data structure is lost; hierarchical spatial information is lost.
- Small object information can not be reconstructed (assuming four pooling layer is any object information is smaller than 2 ^ 4 = 16 pixel theoretically will not be rebuilt.)
In the presence of such a problem, the problem has been in a semantic segmentation of the bottleneck is no longer significantly improve the accuracy and dilated convolution of good design is to avoid these problems.