From 54f1908054e27f3573a59a5fefa725b9ef77996a Mon Sep 17 00:00:00 2001
From: Yuren Cong <47991543+yrcong@users.noreply.github.com>
Date: Sat, 20 Nov 2021 23:47:31 +0100
Subject: [PATCH] Update README.md

---
 README.md | 37 +++++++++++++++++++++++++++++++++++--
 1 file changed, 35 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index f9864b9..2f3b205 100644
--- a/README.md
+++ b/README.md
@@ -1,2 +1,35 @@
-# RelTR
-RelTR: End-to-End Scene Graph Generation with Transformers
+# RelTR: End-to-End Scene Graph Generation with Transformers
+
+We provide the inference code which is entirely implemented in Pytorch in the supplementary. The code of training and evaluation will be released after publication. 
+
+## Installation
+It is very easy to configure the RelTR environment. We strongly recommend running the code on Linux system. Python=3.6, pytorch=1.6 and matplotlib are used in our code.
+Please configure the environment as follows:
+```
+# create a conda environment 
+conda create -n reltr python=3.6
+conda activate reltr
+
+# install packages
+conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1 -c pytorch
+conda install matplotlib
+```
+# Pretrained Model
+Please download the anonymous version of the pretrained RelTR model on the Visual Genome dataset with the link:
+https://drive.google.com/file/d/1id6oD_iwiNDD6HyCn2ORgRTIKkPD3tUD/view
+
+and put it under 
+```
+models/checkpoint0149.pth
+```
+## Usage
+You can infer an image with the command:
+```
+python inference.py --img_path $IMAGE_PATH
+```
+We provide 5 images from VG dataset and 1 image from internet (Please use it only for testing, we do not own the copyright). You can infer the first VG image with:
+```
+python inference.py --img_path images/vg1.jpg
+```
+The result is as follows. The 1st/2nd row is the subject/object attention heat map while the 3rd row shows predicted triplets. Only top-10 confident predictions (and scores of subject, object and predicate>0.3) are shown. For clear demonstration, the attention heatmaps are not overlapped on the predictions as we did in the paper.
+![GitHub Logo](vg1_pred.png)