diff --git a/README.md b/README.md
index 97122ee84db310ddd9d4a9fc06eb5b086ee5690c..69e4879539c16547fb3ceddfd22ae259486ec1e2 100644
--- a/README.md
+++ b/README.md
@@ -10,6 +10,24 @@ Re3 is a real-time recurrent regression tracker. It offers accuracy and robustne
 5. [CUDA (Strongly Recommended)](https://developer.nvidia.com/cuda-downloads).
 6. [cuDNN (Recommended)](https://developer.nvidia.com/cudnn).
 
+## First Time Setup:
+```bash
+git clone git@gitlab.cs.washington.edu:xkcd/re3-tensorflow.git
+cd re3-tensorflow
+sudo apt-get install python-virtualenv
+virtualenv venv --system-site-packages
+source venv/bin/activate
+pip install -r requirements.txt
+```
+### Enter the virtualenv in a later session to use the installed libraries.
+```bash
+source venv/bin/activate
+```
+### To exit the virtualenv
+```bash
+deactivate
+```
+
 ## Model:
 The model weights we used in our paper were ported from Caffe to Tensorflow. Because they deal with padding differently, results are slightly different, but the model still works well.
 To use it, download it [here](https://goo.gl/NWGXGM), extract the tar, and place in your log directory (mine is "logs").
diff --git a/requirements.txt b/requirements.txt
index 4552bad96e209231a86074187b7569ce15de29c9..4e3b2450e2f79c92198007ea983e8e7b3df5d100 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,3 +1,3 @@
-numpy
-opencv-python
+numpy==1.13.1
+opencv-python==3.3.0.10
 tensorflow-gpu==1.2.0