Revisiting Binary Local Image Description for Resource Limited Devices

📢IMPORTANT📢 BAD is now available in OpenCV as cv::xfeatures2d::TEBLID

This repository contains the source code of BAD (renamed as TEBLID after its inclusion in OpenCV contrib) and HashSIFT descriptors presented in “Revisiting Binary Local Image Description for Resource Limited Devices”. When accuracy and efficiency are both important, the descriptors in this repository offer the perfect trade-off for real-time applications and resource limited devices like smartphones, robots or drones.

Dependencies

The code depends on OpenCV 4.

# Install dependencies (Ubuntu 18.04)
sudo apt-get install -y build-essential cmake git libgtk2.0-dev pkg-config libavcodec-dev libavformat-dev libswscale-dev
# Download source code
git clone https://github.com/opencv/opencv.git --branch 4.5.2 --depth 1
# Create build directory
cd opencv && mkdir build && cd build
# Generate makefiles, compile and install
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local ..
make -j
sudo make install

Compile and Run

We provide the pre-trained execution code of BAD and HashSIFT descriptors.
The code contains two demos. The first one estimates the fundamental matrix between
two images of the Machine Hall 05 from EuRoC MAV Dataset.
The demo detects feature points using ORB detector (FAST + Harris score) and describes using BAD.
We draw the RANSAC inliers with less than 3px of epipolar error.
The code can be compiled with Cmake:

mkdir build && cd build
cmake .. && make
./stereo_demo [hashsift]

We also show a second demo that registers a pair of planar images.

./homography_demo [hashsift]

The result for the provided images should be several imshows and something like this in the standard output:

*************** Homography estimation demo ***************
Detected features: 1000
Matched features:   64
Inliers percentage:  6.4%

If ORB descriptor is used instead of BAD, only 2.4% of inliers are obtained.

References

If you use this code, you must cite our Robotics and Automation Letters paper:

@ARTICLE{9521740,
  author={Su\'arez, Iago and Buenaposada, Jos\'e M. and Baumela, Luis},
  journal={IEEE Robotics and Automation Letters}, 
  title={Revisiting Binary Local Image Description for Resource Limited Devices}, 
  year={2021},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/LRA.2021.3107024}}

Contact and Licence

We provide a free pre-trained version of the execution code. Full execution and training code can be obtained under license, if you are interested please contact us:

• Iago Suárez ( [email protected] ) for technical issues.
• Miguel Ángel Orellana Sainz ( [email protected] ) for commercial issues.

This software was developed by The Graffter S.L. in collaboration with the PCR lab of the Universidad Politécnica de Madrid.