As part of our commitment to both open source and reproducible research, BICV has started utilising Docker to create development environments for research into machine learning and computer vision. These environments consist of combinations of operating system + libraries + application environments, where OS/library dependencies are largely resolved. From a practical perspective this reduces the headache of compilation problems that many users struggle with daily. From a scientific perspective it allows code to be run in reproducible environments.
A Dockerfile can automate the building of Docker images. By making these publicly available on our Bitbucket repository, the creation of Docker images is not only transparent, but allows for further Docker images to be built on top of existing ones.
This code implements dense descriptors as described in the ECCV paper “Spatio-chromatic Opponent Features” by Alexiou and Bharath. The code is written in Matlab, but this release includes support for CUDA-acceleration with an appropriate GPU. A pre-print of the paper is available for download from the PDF link below:
As promised, the Small Hand-held Object Recognition Test has been expanded from 30 to 100 categories, including a wider range of groceries, toiletries and other widely available products. This expansion will allow better generalisations of the evaluated algorithms.
The expansion can be browsed from here under the TRAINING-100 directory.
More uploads will follow over the summer: the evaluation code will finally be released, and a single download link will be created for the dataset.