sssemil / challenge2017

HackaTUM2017 challenge by Rohde & Schwarz

Intelligent Logo Detection {#mainpage}

HackaTUM2017 challenge by Rohde & Schwarz

Want a job? Have a look at our job offerings in broadcast and media!

Take our Challenge

Code software to analyze video frames. We provide a rich dataset consisting of 50000 categorized frames from various German TV stations. Use AI and computer vision to automatically detect or classify station logos or excite us by finding an alternative usage for the data.

Discuss on Slack #rohde-schwarz.

Goals

• Detect logos
• Classify logos
• Find exciting alternative usages for the data

Explain to us how to use your code on another dataset. No python interface required.

Documentation

A .html documentation for this repo can be generated by using doxygen.

sudo apt-get install doxygen graphviz
cd doc
doxygen Doxyfile
xdg-open html/index.html

Dataset

Grab an USB stick with the data from our booth.

About the dataset

The dataset consists of video frames recorded from various German free-to-air programs. Every 15 minutes a recording of 9 images was triggered. Additionally, the images were manually grouped depending on sender logos showing up in the frames. This is reflected in the folder structure. There is a file named metadata.txt in each subfolder which describes the available logos.

Filename string

The images are named according to the following format:

<serviceID>_<date>_<time>_[0-8].jpg

• serviceID is a unique id for each service
• date and time specify the capture timestamp

Format of metadata.txt

The file named metadata.txt in each subfolder describes the logos. For each logo, there is one line in the file.

<logo category>,<xPox>,<yPos>,<xSize>,<ySize>

The logo category is a string naming the logo. The position and the size are measured in pixels. xPox, yPos, xSize and ySize describe the bounding box for a logo.

Python library

A python library is provided to read the dataset. Feel free to use any other programming language to work with our data.

Setup

The provided python scripts require numpy and scikit-image. Install on Ubuntu 16.04.

sudo apt-get install python3 python3-pip python3-numpy
pip3 install -U scikit-image

Scripts

Several scripts demonstrate how to use the library.

• dataset_info.py

Illustrate images of the three difficulty levels.

• show_dataset.py
• show_corner_dataset.py
• show_simpledataset.py

Other software that might be useful

• TensorFlow See the TensorFlow webpage for installing it in a virtualized python environment.

pip3 install -U tensorflow

• OpenCV

pip3 install -U python-opencv