GPGPU Project: object detection

A) Goal of the project

The goal of the project is to implement a simple object detector in CUDA.

B) Our group

Name	Email (@epita.fr)	Github account
Corentin Duchêne	corentin.duchene	Nigiva
Théo Perinet	theo.perinet	TheoPeri
Hao Ye	hao.ye	Enjoyshi
Arthur Fan	arthur.fan	arthur942

C) How to run the project

To run the project, it is necessary to have installed :

• Cmake
• CUDA
• Python 3.8.12
• python3-pip
• virtualenv

We use Poetry and Pyenv to generate the python virtual environment but this is not necessary to run the project (it is just recommended).

Run the following commands to import and build the project :

• Git clone the project
• Run python -m venv .venv in the project folder. Be careful to choose the right version of python, for example python3.8 -m venv .venv
• Run source ./.venv/bin/activate to activate the virtual environment
• Run pip install -r requirements.txt
• Run cmake .. in the build/ folder (cd build/)
• Run make all -j in the build/ folder

ℹ️ Note : To build the project in debug mode, you should use...

cmake -DCMAKE_BUILD_TYPE=Debug ..

ℹ️ Note 2 : When building the project, Cmake takes care of downloading an example video Aled.mp4 by itself into the build/data folder.

D) How to use the project ?

D.1) To convert a video to an image folder

To convert a video to an image folder, you can use convert_mp4_to_png.py.

By simply running

python convert_mp4_to_png.py

the video build/data/Aled.mp4 will be converted into a folder of images contained in build/data/export/. This folder is deleted on each re-generation.

If you want to choose the video to be converted or the destination folder for the images, you can read the documentation by running :

python convert_mp4_to_png.py -h

D.2) To run object detection on a series of images

./build/detect build/data/export/reference.png $(ls build/data/export/input-*.png)

D.3) To display the boxes in a video

It is possible to generate a video with the detected objects framed by boxes from the JSON given in STDIN and an image folder.

We provide a JSON example boxes_stdout_example.json to debug our python program. So you can see the result by running the following commands:

• python convert_mp4_to_png.py to generate the image folder from the video supplied by default with the project
• cat boxes_stdout_example.json | python generate_video.py

Normally, you should see :

• 2 green boxes
  • one starts at the top left of the image
  • the other also starts at the top left but shifted 100px to the right and 200px to the bottom.
• The top left box already has a height of 100px at the beginning and increases in height and width linearly.
• The other box has a size of zero at the start but grows faster in width than in height.

To generate the video normally from the STDOUT of the binary, you can therefore run the following command :

./build/detect build/data/export/reference.png $(ls build/data/export/input-*.png) | python generate_video.py

If you want to choose the image folder or the path of the exported video, you can read the documentation by running :

python generate_video.py -h

D.4) To run the benchmark

./build/bench