Fast human action recognition by Local Part Model.
We proposed a novel Local Part Model (LPM) for fast human action recognition. To account for out-of-ordering problem of bag-of-features approach, our local part model includes both a coarse root ST patch covering local event-content statistics and finer overlapping part ST patches integrating local structure and temporal relations. We introduced the idea of using very high sampling density for efficient and accurate action recognition by applying random sampling over LPM and through using integral video. For more information, check out publications [1, 2, 3]. Our code can be found [here]. If you use our code, please cite our works .
We have made a substantial improvement over our CVPR2013 publication. The updated results are shown in the table. In this new improvement, we used separate root and part channels (check ), one-against-all multi-class SVM, a new efficient decriptor and a state-of-the-art optical flow method.
The implementation details and codes for these recent results are from the PhD works to be published soon.