RPF

Official Code Repository for the paper Incorporating Relative Object Positioning for Image Captioning

This Repository will contain code for creating Relative Positioned Features for images and creation and evaluation of the AoANet+VC+RPF model.

All instructions assume you are running from RPF directory.

Contents

• Installation
• Datasets and Features

Installation

Install conda

We recommend installing conda to setup a separate environment as this involves specific versions of certain libraries.

Create Environment

For Linux and Mac

Open a terminal window in RPF directory.

sh src/setup/install.sh "rpf"

This will create an environment called rpf(can be changed by replacing rpf with desired name) for this project and will install the required libraries for the project.

For Windows

Open a command prompt window in RPF directory.

conda create --name rpf
conda activate rpf

This will create an environment called rpf(can be changed by replacing rpf with desired name) for this project.

python src/setup/install.py

Datasets and Features

We used the Karpathy split of the MS-COCO 2014 Dataset which can be downloaded from here or by running download-coco.py by running the below code, but we recommend running the corresponding cell in ipython notebook (RPF_train_colab.ipynb) if using colab. This will download the fll dataset but the code will use only the Karpathy split.

python3 src/setup/download-coco.py