Deep Semantic Matching with Foreground Detection and Cycle-Consistency

Pytorch implementation of our method for weakly supervised semantic matching.

Contact: Yun-Chun Chen (ycchen918 at citi dot sinica dot edu dot tw)

Please cite our paper if you find it useful for your research.

@inproceedings{Chen_WeakMatchNet_2018,
  author = {Y.-C. Chen and P.-H. Huang and L.-Y. Yu and J.-B. Huang and M.-H. Yang and Y.-Y. Lin},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  title = {Deep Semantic Matching with Foreground Detection and Cycle-Consistency},
  year = {2018}
}

Installation

Install PyTorch
Clone this repo

git clone https://github.com/YunChunChen/WeakMatchNet
cd WeakMatchNet

Dataset

Please use the code to download the datasets and put it under the data/ folder.
Please download the pre-trained model for training here and put it under the trained_models/resnet101/ folder.
Evaluation command

sh eval.sh

Training command

sh train.sh

Related Implementation and Dataset

Rocco et al. Convolutional Neural Network Architecture for Geometric Matching. In CVPR, 2017. [project] [paper] [code]
Rocco et al. End-to-End Weakly-Supervised Semantic Alignment. In CVPR 2018. [project] [paper] [code]

Acknowledgment

This code is heavily borrowed from weakalign.

Note

The model and code are available for non-commercial research purposes only.

Nov 2018: code released!