Implementation of the method proposed in the paper:
Outfit Compatibility Prediction and Diagnosis with Multi-Layered Comparison Network
by Xin Wang, Bo Wu, Yueqi Zhong Published at ACM MM 2019 in Nice, France.
- mcn: Main program source code
- data: Polyvore-T datasets based on Polyvore.
- baselines: Compared baselines in our experiment
- exp: Experiment details, scripts and results etc.
Ubuntu 16.04, NVIDIA GTX 1080Ti (for batch size 16), python >= 3.5.2
torch>=0.4.1
torchvision
networkx
-
Download the original Polyvore dataset, then unzip the file and put the
image
directory intodata
folders (or you can create a soft link for it). -
Train
cd mcn python train.py
-
Evaluate
python evaluate.py
-
Visualize outfit diagnosis
cd exp python diagnosis.py
-
Automatically revise outfit
python revision.py
Pretrained model weights can be found in the links. The train, validation and test split is provided in data.
AUC | FITB | |
---|---|---|
Pooling | 88.35 | 57.28 |
Concatenation | 83.40 | 52.91 |
Self-attention | 79.65 | 48.60 |
BiLSTM | 74.82 | 46.02 |
CSN | 84.90 | 57.06 |
Ours | 91.90 | 64.35 |
Please cite our paper if you use this code in your own work:
@inproceedings{wang2019diagnosis,
title={Outfit Compatibility Prediction and Diagnosis with Multi-Layered Comparison Network},
author={Xin Wang, Bo Wu and Yueqi Zhong},
booktitle={ACM International Conference on Multimedia},
year={2019}
}