Paper Learning
The learning of "Deep Saliency with Encoded Low level Distance Map and High Level Features" (2016, CVPR, pages 660-668)
1.1.1 Theories
1) The performance using deep learning to obtain high level features to detect salient regions outperform previous works that utilized only low level features.
2) High level features are good to evaluate objectness in an image, they are relatively weak in for determining precise localization. (Multiple levels of convolutional and polling layers “blur” the object boundaries, and it was still very hard to differentiate salient regions from their adjacent non-salient regions because their feature distances were not directly encoded.)
3) The hand-crafted features can provide complementary information to enhance performance of saliency detection that utilizes only high level features.
1.1.2 Model overview
1) Getting the high level feature from the output of the conv5 3 layer of the VGG16 model.
2) Getting the low level feature by computing the ELD-map from the intial feature distance map for each query region.
3) Concatenate our ELD-map and the output of the last convolutional layer from the VGG16 to form a new feature vector, then evaluated by a multi-level fully connected neural network classifier (cross entropy loss for softmax classifier).
1.1.3Worth learning highlights
1) The method to get initial low level features.distance maps
 |
Throuth SLIC algorithm for superpixels segementation, use colors (average colors in RGB, LAB, and HSV spaces, and their local color histograms), textures (Gabor filter responses averaged over pixels in each region), and locations (center location of a superpixel) as low level features.
 |
The superpixels as S ={r1,...,rM}, and the grid cells of the regular grid as G = {c11,c12,...,cNN}, N = 23.K=54(depth) (Every feature forms one layer feature map.)
fcij≔f(rc*), where rc*=argmaxrc
1.1.4 Problem
How to get query region? This question is not mentioned.