We proudly announce the release of ARTOS 2.0 which comes with a few new features as well as a completely refactored feature extraction framework that allows for the usage of arbitrary feature extractors (e.g. extracted from a CNN).

Some of the most prominent features and changes are:

• A Model Evaluation API for libartos and PyARTOS, including GUI support for evaluating models and plotting recall-precision graphs.
• Completely refactored feature extraction: Any existing dependencies on HOG specifics have been removed from ARTOS and FFLD, which are not strictly separated any more. A new feature extraction framework provides an abstraction layer which should make it easy to implement arbitrary feature extractors. The GUI has been extended so that the feature extractor and its parameters can be changed at run-time in a convenient way.
• ARTOS v2 ships with the new CaffeFeatureExtractor as an alternative to HOG. It uses the Caffe library to extract image features from the layer of a CNN which can then be used with our fast linear detector. Please refer to README.md and the documentation of CaffeFeatureExtractor for instructions on how to use this new feature extractor.
• GUI support for alternative image repository drivers (e.g. plain directories instead of ImageNet).
• Optimized code which makes ARTOS even faster.

A more complete list of changes can be found in the CHANGELOG.md file.

However, the introduction of a new major version means that there are some backwards-incompatible changes. The following two probably are most important:

• The format of model files and background statistics files created by ARTOS v2 differs from the file format used by previous versions, since those files now contain information about the feature extractor. For the sake of compatibility, ARTOS can still read files in the old format under the assumption that they have been created with default HOG features, but it will only create files in the new format.
• The padding parameter has been removed from the function create_detector of the C API. The necessary amount of padding is now determined by the feature extractor automatically.
• Building ARTOS new requires a C++11-compliant compiler.
• ARTOS-related CMake options have been renamed and prefixed with ARTOS_.

A more extensive list of changes which affect existing code and applications can be found in the CHANGELOG.md file.

This is mainly a bugfix release, which also adds the option to disable caching of positive images during training if the available RAM is very limited.

Key features of this release:

• Possibility to use a plain directory structure for your images and annotation files instead of those tar archive used by ImageNet (refer to README.md for details).
• Added DPMDetection::detectMax(), which yields just the highest scoring detection on a given image (may be useful for classification-like tasks).
• Added an abstraction layer to the model learning process: The new abstract base class ModelLearnerBase is agnostic of the concrete learning method, which may be the WHO method implemented by ModelLearner as well as any other linear classifier.

Besides that, some minor bugs and ugly memory leaks have been fixed.
Please refer to CHANGELOG.md for details and more changes.

This release comes with improved background statistics computed from 32k ImageNet samples. We've also added a novel method for fast computation of such statistics, which leverages the Fourier transform.

A number of minor bugs has been fixed too. Please refer to CHANGELOG.md for details.