vtranse/STA Tensorflow

visual translation embedding network for visual relation detection, CVPR 2017, tensorflow

Shuffle-Then-Assemble: Learning Object-Agnostic Visual Relationship Features, ECCV, tensorflow

Installation

1. Install ipython, if you do not have ipython, you can install this tool (strongly recommended: https://ipython.org/install.html)

pip install ipython

2. Install TensorFlow v1.3.0 or newer type.

pip install tensorflow-gpu==1.3.0

3.Download this repository or clone with Git

git clone https://github.com/yangxuntu/vtranse.git

3. Install easydict

pip install easydict

Training and Testing Vtranse

1. Download dataset (VRD dataset is used as example)

a). Download the dataset form https://share.weiyun.com/55KK78Y, and the file is named as 'sg_dataset.zip'.

b). Use the following commend to unzip the downloaded data:

unzip sg_dataset.zip -d sg_dataset

c).In the path where you put vtranse folder, use the following commend to make a new folder 'dataset/VRD':

mkdir -p ~/dataset/VRD/json_dataset
mkdir -p ~/dataset/VRD/sg_dataset

d). Move the files in sg_dataset into the created dataset, by using the following commends:

mv sg_dataset/annotations_test.json dataset/VRD/json_dataset
mv sg_dataset/annotations_train.json dataset/VRD/json_dataset
mv sg_dataset/sg_test_images dataset/VRD/sg_dataset
mv sg_dataset/sg_train_images dataset/VRD/sg_dataset

e). Change the root path in file 'vtranse/model/config.py': open this file and find the term '__C.DIR' which is named as '/home/yangxu/rd' to suitable path where you put this vtrase folder.

f). Pre-process the VRD dataset to the vrd_roidb.npz which can be used to train the network. Open ipython using following commend:

ipython

And then use following commend to pre-process data in vrd folder:

run process/vrd_pred_process.py

After runing this file, you will find that there is one 'vrd_roidb.npz' file in the foloder 'vtranse/input'

2. Training

a). Download pre-trained model of faster-rcnn on VRD dataset from https://share.weiyun.com/5skGi9N, and the file names are 'vrd_vgg_pretrained.ckpt.data-00000-of-00001', 'vrd_vgg_pretrained.ckpt.index', 'vrd_vgg_pretrained.ckpt.meta' and 'vrd_vgg_pretrained.ckpt.pkl'. After downloading them, using the following commend to move them into the 'vtranse/pre_trained' file:

mv vrd_vgg_pretrained.ckpt.data-00000-of-00001 vtranse/pretrained_para
mv vrd_vgg_pretrained.ckpt.index vtranse/pretrained_para
mv vrd_vgg_pretrained.ckpt.meta vtranse/pretrained_para
mv vrd_vgg_pretrained.ckpt.pkl vtranse/pretrained_para

b). Create a folder which is used to save the trained results

mkdir -p ~vtranse/pred_para/vrd_vgg

c). After downloading and moving files to suitable folder, using 'vtranse/train_file/train_vrd_vgg.py' to train vtranse network on VRD dataset.

ipython
run train_file/train_vrd_vgg.py

d). When training, you can see the results like that:

t: 100.0, rd_loss: 4.83309404731, acc: 0.0980000074953
t: 200.0, rd_loss: 3.81237616211, acc: 0.263000019006
t: 300.0, rd_loss: 3.51845422685, acc: 0.290333356783
t: 400.0, rd_loss: 3.31810754955, acc: 0.292666691653
t: 500.0, rd_loss: 3.48527273357, acc: 0.277666689083
t: 600.0, rd_loss: 3.06100189149, acc: 0.340666691475
t: 700.0, rd_loss: 3.02625158072, acc: 0.334666692317
t: 800.0, rd_loss: 3.06034492403, acc: 0.330333357863
t: 900.0, rd_loss: 3.16739703059, acc: 0.322666690871
...

3. Testing

a). After training vtranse, you will find files like 'vrd_vgg0001.ckpt' in the 'vtranse/pred_para/vrd_vgg' folder. And then you can test your trained model

b). Open the file 'vtranse/test_file/test_vrd_vgg_pred.py' and change the variable 'model_path' to the suitable pretrained model's name.

c). Create a folder to save the result of detected relationships using the following commend:

mkdir -p ~vtranse/pred_res

d). After changing the name of your model, using following commend to get the relationship detection results:

ipython
run test_file/test_vrd_vgg_pred.py

e). After testing, you can run the file 'vtranse/test_file/eva_vrd_vgg_pred.py' to evaluate your detected result:

ipython
run test_file/eva_vrd_vgg_pred.py

VG dataset

1). Download VG dataset. This dataset can be downloaded from their offical website: https://visualgenome.org/. After downloading these files, you should using the following commend to put these images into the folder 'dataset/VG/images/VG_100K'

mkdir -p ~dataset/VG/images/VG_100K
mv images/VG100K dataset/VG/images/VG_100K
mv images/VG100K dataset/VG/images/VG_100K

2). Download training/testing split Since this dataset is so noisy, and I use one filtered type which is provided by https://drive.google.com/file/d/1C6MDiqWQupMrPOgk4T12zWiAJAZmY1aa/view?usp=drive_web, you can download the split form this link. After downloading this file, you can use the following commend to pre-process the vg dataset

mkdir -p ~dataset/VG/imdb
mv vg1_2_meta.h5 dataset/VG/imdb
ipython
run process/vg_pred_process.py

3). Training, Testing and Evaluation After pre-processing Vg dataset, you can using similar process like VRD dataset to train, test and evaluate your model by using following commends:

ipython
run train_file/train_vg_vgg.py

ipython
run test_file/test_vg_vgg_pred.py

ipython
run test_file/eva_vg_vgg_pred.py

Citation:

@inproceedings{Zhang_2017_CVPR,
  author    = {Hanwang Zhang, Zawlin Kyaw, Shih-Fu Chang, Tat-Seng Chua},
  title     = {Visual Translation Embedding Network for Visual Relation Detection},
  booktitle = {CVPR},
  year      = {2017},
}

Results of VRD (R100)

	predicate	phrase	relation
published result	44.76	22.42	15.20
implemented result	46.48	24.32	16.27

Results of VG (R100)

	predicate	phrase	relation
published result	62.87	10.45	6.04
implemented result	61.70	13.62	11.62

References:

1. VRD project: https://cs.stanford.edu/people/ranjaykrishna/vrd/

2. Visual Genome https://visualgenome.org

3. Vtranse Caffe Type: https://github.com/zawlin/cvpr17_vtranse

4. The faster rcnn code which I used to train the detection part in this file: https://github.com/endernewton/tf-faster-rcnn

Contact

1. If you have any problems of this programming, you can eamil to [email protected].