EVA (Exploratory Video Analytics)

What is EVA?

EVA is a visual data management system (think MySQL for videos). It supports a declarative language similar to SQL and a wide range of commonly used computer vision models.

What does EVA do?

• EVA enables querying of visual data in user facing applications by providing a simple SQL-like interface for a wide range of commonly used computer vision models.

• EVA improves throughput by introducing sampling, filtering, and caching techniques.

• EVA improves accuracy by introducing state-of-the-art model specialization and selection algorithms.

Table of Contents

• Installation
• Client Testing
• Docker
• Development
• Architecture

Installation

Installation of EVA involves setting a virtual environment using miniconda and configuring git hooks.

1. Clone the repository

git clone https://github.com/georgia-tech-db/eva.git

2. Install the dependencies.

sh script/install/before_install.sh
export PATH="$HOME/miniconda/bin:$PATH"
sh script/install/install.sh

3. Connect mysql user root with normal account and no password

sudo mysql -u root
> SELECT User,Host FROM mysql.user;
> DROP USER 'root'@'localhost';
> CREATE USER 'root'@'%' IDENTIFIED BY '';
> GRANT ALL PRIVILEGES ON *.* TO 'root'@'%' WITH GRANT OPTION;
> FLUSH PRIVILEGES;
mysql -u root

refer to askubuntu

4. Install docker and docker-compose. Please refer to official doc.

Client Testing

1. Set up the server and client

• Launch EVA database Server: python eva.py

• Launch CLI: python eva_cmd_client.py

2. Run the UPLOAD command in the client terminal:

UPLOAD INFILE 'data/ua_detrac/ua_detrac.mp4' PATH 'test_video.mp4';

3. Run the LOAD command in the client terminal: (may take a while)

LOAD DATA INFILE 'test_video.mp4' INTO MyVideo;

4. Below is a basic query that should work on the client

SELECT id, data FROM MyVideo WHERE id < 5;

Docker

1. Standup EVA testing for CPU/GPU hardware.

docker-compose -f docker-compose.yml eva-test-[cpu/gpu] up

Development

1. Install git hooks in your .git/ directory. [optional, but recommended]

conda activate eva
pre-commit install

2. Ensure that all the unit test cases (including the ones you have added) run succesfully and the coding style conventions are followed.

bash script/test/test.sh

Quickstart Tutorial

Configure GPU (Recommended)

1. If your workstation has a GPU, you need to first set it up and configure it. You can run the following command first to check your hardware capabilities.

   ubuntu-drivers devices
   

   If you do have an NVIDIA GPU, and its not been configured yet, follow all the steps in this link carefully. https://towardsdatascience.com/deep-learning-gpu-installation-on-ubuntu-18-4-9b12230a1d31.

   Some pointers:

   • When installing NVIDIA drivers, check the correct driver version for your GPU to avoid compatibiility issues.
   • When installing cuDNN, you will have to create an account. Make sure you get the correct deb files for your OS and architecture.

2. You can run the following code in a jupyter instance to verify your GPU is working well along with PyTorch.

   import torch
   device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
   print(device)
   

   Output of cuda:0 indicates the presence of a GPU. (Note: 0 indicates the index of the GPU in system. Incase you have multiple GPUs, the index needs to be accordingly changed)

3. Now configure the executor section in eva.yml as follows:

   gpus: {'127.0.1.1': [0]}
   

   127.0.1.1 is the loopback address on which the eva server is started. 0 refers to the GPU index to be used.

Sample Notebook

1. Open a terminal instance and start the server:

   python eva.py
   

2. Open another terminal instance. Start a jupyter lab/notebook instance, and navigate to tutorials/object_detection.ipynb

3. You might have to install ipywidgets to visualize the input video and output. Follow steps in https://ipywidgets.readthedocs.io/en/latest/user_install.html as per your jupyter environment.

4. Run each cell one by one. Each cell is self-explanatory. If everything has been configured correctly you should be able to see a ipywidgets Video instance with the bounding boxes output of the executed query.

Architecture

EVA consists of four core components:

• EVAQL Query Parser
• Query Optimizer
• Query Execution Engine (Filters + Deep Learning Models)
• Distributed Storage Engine

Contributing

To file a bug or request a feature, please file a GitHub issue. Pull requests are welcome.

Contributors

See the people page for the full listing of contributors.

License

Copyright (c) 2018-2020 Georgia Tech Database Group Licensed under the Apache License.