# Filter Spatial Channel Weighting

+ Author : 利醬
+ Date : 2021/08/13

## Source
International Journal of Pattern Recognition and Artificial Intelligence
2021 -
Filtering deep convolutional features for image retrieval

## PyTorch Implementation
<https://github.com/leonIceHot/Computer-Vision/tree/master/Filtering-Spatial-Channel-Weighting>

Project Architecture

Filtering on spatial channel weighting (FSCW)
-----
+ 

+ Preprocessing
+ Model
+ Filter
+ Spatial
+ Channel Weighting
+ Image Representation

### Preprocssing
+ Dataset :
+ Oxford5K
+ The Oxford Buildings Dataset
+ <https://www.robots.ox.ac.uk/~vgg/data/oxbuildings/>

+ Paris6K 、 Paris106K

+ Holdout validation
+ Total : 5063 images 
+ Train : 0.7 (3545)
+ Validation : 0.3 (1518)
+ Test : 0.5 (2531)

### Model
Pytorch implementation of cnn network
+ <https://github.com/shanglianlm0525/PyTorch-Networks>
+ VGG
+ 
+ 

+ ResNet
+ 

### Filtering the feature maps
We extract the feature maps in the pool5 layer of VGG16 model,
Whereas we extract the feature maps in the res5c_relu layer of the ResNet101 model

+ 𝑋 ∈ 𝑅(𝑊×𝐻×𝑁) be the 3-dim feature tensor extracted
+ N = 512 feature maps
+ W: Width of a feature map
+ H: Height of a feature map




### Spatial Weighting




### Channel Weight

### Image Representation

1. Sum-Pooling

2. Global representation vector

3. FSCW descriptor
+ PCA Whitening
+ Theorem
+ <http://ufldl.stanford.edu/tutorial/unsupervised/PCAWhitening/>
+ <http://ufldl.stanford.edu/tutorial/unsupervised/ExercisePCAWhitening/>
+ <https://www.twblogs.net/a/5b7e231c2b717768385582bf>
+ Implementation
+ sklearn.decomposition.PCA
+ <https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html>

+ L2 Normalization

## Experiment

### Query expansion (QE)