Atrous algorithm: deeplab

This algorithm first appeared in deep learning is Deeplab for semantic segmentation.

In Review of Deep learning [nature14539]

An image, for example, comes in the form of an array of pixel values, and the learned features in the first layer of representation typically represent the presence or absence of edges at particular orientations and locations in the image. The second layer typically detects motifs by spotting particular arrangements of edges, regardless of small variations in the edge positions. The third layer may assemble motifs into larger combinations that correspond to parts of familiar objects, and subsequent layers would detect objects as combinations of these parts.

Then there is a trade-off between spatial information and semantic information. In shallow layers, there are more spatial info but less semantic info. In deeper layers, we get more semantic info but we lose the location. However,these high-level CV tasks require both of them the more complete the better.

In FCN, the last layer’s feature map only obtain 16*16 size, then bilinear interpolation. It is better off using a denser feature map to achieve better spatial performance. Then deeplab transfer the atrous algorithm to this elegant transfer learning which is widely used in object detection, semantic segmentation, etc. In this paper:

We can implement this more efficiently by keeping the filters intact and instead sparsely sample the feature maps on which they are applied on using an input stride of 2 or 4 pixel, respectively.

Aiming to achieve denser feature map but without retrain pipeline, we have to keep the filters intact, namely, keep the receptive field intact..
So
1. Only to fine-tune the “head layers” added
2. Utilize the model’s filter downloaded sufficiently.

---

Atrous algorithm solve the problom that how to keep the filters
Here is an 1 dimension example.

a. original VGG with poor feature maps.
b. denser one but change the receptive field.
c. atrous algorithm.

快考试了得练一下英语，如果英语或者算法理解有错，感谢交流指正