Beyond Joints: Learning Representations from Primitive Geometries for Skeleton-based Action Recognition and Detection
Hongsong Wang, Liang Wang. Beyond Joints: Learning Representations from Primitive Geometries for Skeleton-based Action Recognition and Detection. IEEE Transactions on Image Processing (TIP 2018), vol. 27, no. 9, pp. 4382-4394. [Published article]
Our codebase is based on Keras with Theano backend. If you use tensorflow backend, you need modify it in some ways.
Please read this [ntu rgb+d repository] and download the dataset by following this [instruction]. You can only download the skeletons of this dataset.
In the ntu_rgbd.py, we use function save_h5_file_skeleton_list to save skeletons into hdf5 file. If one sample has more than one skeletons, we save one skeleton each. During traning, we sample sub-sequences with length of 100 with an interval of 50, and assign it with the action label. For testing, we average the predicted scores of this subsequences. If the test sample has multiple skeletons, we average scores of sub-sequences of all the skeletons. The details are in the function group_person_list and the class evaluateVal in the deep_rnn.py.
We follow the experimental setup of this paper, Cooccurrence Feature Learning for Skeleton based Action Recognition using Regularized Deep LSTM Networks, AAAI 2016. The processed skeleton data can be download from the the [website] (search for key words: CMU subset ). Here, we use 4-fold cross validation files for CMU dataset, not the three fold cross validation files for CMU subset.
The dataset [PKU-MMD repository]. You can only download the skeletons. The main data processing is the function load_seq_lst in classify_pertime.py. During training, we only sample sub-squences with length of 200 around the action intervals. It seems this step is a bit important for fast training, when compared with random sampling.
For any question, feel free to contact: Hongsong Wang, hongsongsui@gmail.com