This repository contains implementation for Grounding of Textual Phrases in Images by Reconstruction in ECCV 2016.

Note: Please read the feature representation files in feature and annotation directories before using the code.

Platform: Tensorflow-1.0.1 (python 2.7)
Visual features: We use Faster-RCNN pre-trained on PASCAL 2012 VOC for Flickr30K Entities, and pre-trained on ImageNet for Referit Game. Please put visual features in the feature directory (More details can be seen in the README.md in this directory). (Fine-tuned features can achieve better performance, which are available in this repository).
Sentence features: We encode one-hot vector for each query, as well as the annotation for each query and image pair. Please put the encoded features in the annotation directory (More details are provided in the README.md in this directory).
File list: We generate a file list for each image in the Flickr30K Entities. If you would like to train and test on other dataset (e.g. Referit Game), please follow the similar format in the flickr_train_val.lst and flickr_test.lst.
Hyper parameters: Please check the Config class in the train_supervise.py and train_unsupervise.py.

We implement both supervised and unsupervised scenarios of GroundeR model.

For training, please enter the root folder of GroundeR, then type

$ python train_supervise.py -m [Model Name] -g [GPU ID]

For testing, please enter the root folder of GroundeR, then type

$ python evaluate_supervise.py -m [Model Name] -g [GPU ID] --restore_id [Restore epoch ID]

Make sure the model name entered for evaluation is the same as the model name in training, and the epoch id exists.

The implementation of unsupervised model of GroundeR is a little different from the paper: In Equation 5, original GroundeR adopts a softmax function to calculate attention weights, while we adopt a relu function to generate these weights. We observe a performance drop by using softmax function. To try original GroundeR model, please uncomment line 96 and comment line 97 in model_unsupervise.py.
For training, please enter the root folder of GroundeR, then type

$ python train_unsupervise.py -m [Model Name] -g [GPU ID]

For testing, please enter the root folder of GroundeR, then type

$ python evaluate_unsupervise.py -m [Model Name] -g [GPU ID] --restore_id [Restore epoch ID]

Make sure the model name entered for evaluation is the same as the model name in training, and the epoch id exists.