Awesome
Note
This API is able to load all 4 referring expression datasets, i.e., RefClef, RefCOCO, RefCOCO+ and RefCOCOg. They are with different train/val/test split by UNC, Google and UC Berkeley respectively. We provide all kinds of splits here.
<table width="100%"> <tr> <td><img src="http://bvisionweb1.cs.unc.edu/licheng/referit/refer_example.jpg", alt="Mountain View" width="95%"></td> </tr> </table>Citation
If you used the following three datasets RefClef, RefCOCO and RefCOCO+ that were collected by UNC, please consider cite our EMNLP2014 paper; if you want to compare with our recent results, please check our ECCV2016 paper.
Kazemzadeh, Sahar, et al. "ReferItGame: Referring to Objects in Photographs of Natural Scenes." EMNLP 2014.
Yu, Licheng, et al. "Modeling Context in Referring Expressions." ECCV 2016.
Setup
Run "make" before using the code.
It will generate _mask.c and _mask.so in external/ folder.
These mask-related codes are copied from mscoco API.
Download
Download the cleaned data and extract them into "data" folder
As the webserver is broken (sry about this), please check this Issue for all datasets downloading. Sorry about the trouble!
Prepare Images:
Besides, add "mscoco" into the data/images folder, which can be from mscoco
COCO's images are used for RefCOCO, RefCOCO+ and refCOCOg.
For RefCLEF, please add saiapr_tc-12 into data/images folder. We extracted the related 19997 images to our cleaned RefCLEF dataset, which is a subset of the original imageCLEF. Download the subset and unzip it to data/images/saiapr_tc-12.
How to use
The "refer.py" is able to load all 4 datasets with different kinds of data split by UNC, Google, UMD and UC Berkeley. Note for RefCOCOg, we suggest use UMD's split which has train/val/test splits and there is no overlap of images between different split.
# locate your own data_root, and choose the dataset_splitBy you want to use
refer = REFER(data_root, dataset='refclef', splitBy='unc')
refer = REFER(data_root, dataset='refclef', splitBy='berkeley') # 2 train and 1 test images missed
refer = REFER(data_root, dataset='refcoco', splitBy='unc')
refer = REFER(data_root, dataset='refcoco', splitBy='google')
refer = REFER(data_root, dataset='refcoco+', splitBy='unc')
refer = REFER(data_root, dataset='refcocog', splitBy='google') # test split not released yet
refer = REFER(data_root, dataset='refcocog', splitBy='umd') # Recommended, including train/val/test