Code for our CVPR 2021 paper "Room-and-Object Aware Knowledge Reasoning for Remote Embodied Referring Expression".

Contributed by Chen Gao*, Jinyu Chen*, Si Liu1†, Luting Wang, Qiong Zhang, Qi Wu

1. Clone this repository.

   git clone https://github.com/alloldman/CKR.git $CKR-root
   

2. Install pytorch==1.3.0

   conda install pytorch=1.3.0 cudatoolkit=9.0 torchvision -c pytorch
   

3. Install the requirements.

   pip install -r requirements.txt
   

1. Download ResNet-152 features for Matterport 3D dataset:

   wget https://www.dropbox.com/s/o57kxh2mn5rkx4o/ResNet-152-imagenet.zip -P img_features/
   unzip ResNet-152-imagenet.zip
   

2. Download the Intermediate data from here. data.zip, cache.zip, img_features.zip, best-ckpt.zip should be unziped. And the KB:data should be download and be unziped under the KB folder.

3. Put these unziped files as the order below:

   CKR
   ├──data
   ├──KB
   │  ├──cache
   |  └─data
   ├──experiments
   │  └──best-ckpt
   └──img_features
   └──ResNet-152-imagenet.tsv 
   

1. Execute the commond below. '0' means using the number 0 GPU.

    bash run.sh train 0

1. Evalution by our rewritten script and select the best checkpoint. An example evalution on REVERIE dataset as follow. You can change the path to evalution your own checkpoint:

   bash run.sh search experiments/best-ckpt/follower_pm_sample2step_imagenet_mean_pooled_1heads_train_iter_9300val_seen_sr_0.547_val_unseen_sr_0.138_ 0
   

Please consider citing this project in your publications if it helps your research. The following is a BibTeX reference. The BibTeX entry requires the url LaTeX package.

@inproceedings{gao2021room,
  title={Room-and-Object Aware Knowledge Reasoning for Remote Embodied Referring Expression},
  author={Gao, Chen and Chen, Jinyu and Liu, Si and Wang, Luting and Zhang, Qiong and Wu, Qi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2021}
}

CKR-nav is released under the MIT license. See LICENSE for additional details.

Some of the codes are built upon REVERIE and babywalk. Thanks them for their great works!