[Paper] [Slide]

This repository contains the code for our paper Dual-Path Convolutional Image-Text Embedding. Thank you for your kindly attention.

What's New: We updated the paper to the second version, adding more illustration about the mechanism of the proposed instance loss.

I have included my Matconvnet in this repo, so you do not need to download it again.You just need to uncomment and modify some lines in gpu_compile.m and run it in Matlab. Try it~ (The code does not support cudnn 6.0. You may just turn off the Enablecudnn or try cudnn5.1)

If you fail in compilation, you may refer to http://www.vlfeat.org/matconvnet/install/

1. Extract wrod2vec weights. Follow the instruction in ./word2vector_matlab;

2. Prepocess the dataset. Follow the instruction in ./dataset. You can choose one dataset to run. Three datasets need different prepocessing. I write the instruction for Flickr30k, MSCOCO and CUHK-PEDES.

3. Download the model pre-trained on ImageNet. And put the model into './data'.

(bash) wget http://www.vlfeat.org/matconvnet/models/imagenet-resnet-50-dag.mat

Alternatively, you may try VGG16 or VGG19.

You may have a different split with me. (Sorry, this is my fault. I used a random split.) Just for a backup, this is the dictionary archive used in the paper.

You may download the three trained models from GoogleDrive.

• For Flickr30k, run train_flickr_word2_1_pool.m for Stage I training.

Run train_flickr_word_Rankloss_shift_hard for Stage II training.

• For MSCOCO, run train_coco_word2_1_pool.m for Stage I training.

Run train_coco_Rankloss_shift_hard.m for Stage II training.

• For CUHK-PEDES, run train_cuhk_word2_1_pool.m for Stage I training.

Run train_cuhk_word_Rankloss_shift for Stage II training.

Select one model and have fun!

• For Flickr30k, run test/extract_pic_feature_word2_plus_52.m and to extract the feature from image and text. Note that you need to change the model path in the code.

• For MSCOCO, run test_coco/extract_pic_feature_word2_plus.m and to extract the feature from image and text. Note that you need to change the model path in the code.

• For CUHK-PEDES, run test_cuhk/extract_pic_feature_word2_plus_52.m and to extract the feature from image and text. Note that you need to change the model path in the code.

• Get word2vec weight

• Data Preparation (Flickr30k)

• Train on Flickr30k

• Test on Flickr30k

• Data Preparation (MSCOCO)

• Train on MSCOCO

• Test on MSCOCO

• Data Preparation (CUHK-PEDES)

• Train on CUHK-PEDES

• Test on CUHK-PEDES

• Run the code on another machine