cvpr17-dvsq

This is the Tensorflow (Version 0.11) implementation of CVPR-17 paper "Deep Visual-Semantic Quantization for Efficient Image Retrieval". The descriptions of files in this directory are listed below:

• net.py: contains the main implementation (network structure, loss function, optimization procedure and etc.) of the proposed approach dvsq.
• net_val.py: contains the implementation of dvsq for evaluation.
• util.py: contains the implementation of Dataset, MAP and ProcessBar.
• train_script.py: gives an example to show how to train dvsq model.
• validation_script.py: gives an example to show how to evaluate the trained quantization model.
• run_dvsq.sh: gives an example to show the full procedure of training and evaluating the proposed approach dvsq.

Data Preparation

In data/nuswide_81/train.txt, we give an example to show how to prepare image training data. In data/nuswide_81/test.txt and data/nuswide_81/database.txt, the list of testing and database images could be processed during predicting procedure. In data/nuswide_81/nuswide_wordvec.txt, we have already prepared the word vectors of the labels extracted by Word2Vec model pretrained on Google News Dataset.

Training Model and Predicting

The bvlc_reference_caffenet is used as the pre-trained model. If the NUS_WIDE dataset and pre-trained caffemodel is prepared, the example can be run with the following command:

"./run_dvsq.sh"

Citation

@inproceedings{conf/cvpr/CaoL0L17,
  author    = {Yue Cao and
               Mingsheng Long and
               Jianmin Wang and
               Shichen Liu},
  title     = {Deep Visual-Semantic Quantization for Efficient Image Retrieval},
  booktitle = {2017 {IEEE} Conference on Computer Vision and Pattern Recognition,
      {CVPR} 2017, Honolulu, Hawaii, USA, July 21-26, 2017}
}