Deep Learning for Computer Vision

Using trajectory for classification

Wang et al. (https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Wang_Action_Recognition_With_2015_CVPR_paper.pdf) used the trajectory of parts of bodies to classify the actions performed. This work combines handcrafted and deep learned features for final predictions. The following is a representation of the classification:

Reproduced from Wang et al.

The handcrafted features are Fisher vector and the features are from CNN. The following image demonstrates the extraction of the trajectories and features maps:

Reproduced from Wang et al.

Both the trajectories and features maps are combined ...