the code of the paper, FashionNet. using keras, on jupyter
using the data set of the FashionNet's subset, in-shop_clothes_retrieval_benchmark
├── data # all input and output data are saved here
│ ├── input -> /data/DeepFashion/in-shop_clothes_retrieval_benchmark # input data
│ └── output # output data
├── src # source code
│ ├── DatasetPreprocessing.ipynb # data set preprocessing code
│ └── FashionNet.ipynb # FashionNet code
└── README.md-
Preprocessing code has been completed (In DatasetPreprocessing.ipynb).
-
Data set reading code of model training has been completed (In FashionNet.ipynb).
-
The code of model structure has been completed (In FashionNet.ipynb).
-
The code for model predicting has not been completed.
FashionNet is used to predict landmarks, category, attribute in clothing images. And its network structure is as follows.
The network structure consists of three parts, blue, red, and green. And they use the same base model, the first 4 layers of VGG16 pretrained on imagenet data set.
The blue part of FashionNet is used to predict the landmarks location and landmark visibility. It's exactly the same as the VGG16 backend network, except for the ouput layer
The red part of FashionNet is used to generate the global features of images. It's the same as the conv4 and first dense layer of VGG16.
The green part of FashionNet is used to generate the local features near landmarks of images.
- pool5_local: the same as ROI pooling. And the ROI is generated from imaages near landmarks
We simply concatenate the green and red part for attribute, category and triplet prediction.
You can download the dataset from All FashionNet Dataset
Don't worry, I've commented on the code in detail. If you encounter any training problem, please feel free to contact me.
Since time limited, the prediction code has not been implemented. Maybe I'll finish it later.
After all, if you want to predict the landmarks or category of the clothing images, just use the blue part. And if you want to predict the attribute or category of the clothing images, just use the red and green part.