Anomaly detection framework (ECCV 2016)
We have developed a framework for anomaly detection in which no training data is required. Simply provide it a set of points, and it will produce a set of anomaly 'ratings', with the most anomalous points producing the highest scores. The algorithm is discriminative, so it has fewer assumptions than model-based approaches and is easily parallelizable.
Paper: poster | pdf | arXiv; also appeared at ECCV 2016.
Code: anomalyframework (Python). Old version (MATLAB).
This code takes .train files (libsvm format) and produces anomaly scores for each feature. The code is highly parallelized, so running on a machine with more CPUs will produce faster results.
At the moment, it is only available for Linux machines with MATLAB installed. If I get enough requests for a Python version, I will make that happen.
Feature generation: videofeatures
This code allows you to generate the same features we used in our experiments, for any video you'd like. These feature files can then be fed into the anomaly detection framework code.
Informative features for anomaly detection (ICML 2016 Workshop on Anomaly Detection)Poster: ICML2016Workshop.pdf
Code: Available upon request (but very easy to implement yourself with eig in either MATLAB or Python!).