analyze

Tools for computing movement/pose similarity time series, clustering

compare_sequences_pairwise

 compare_sequences_pairwise (seq1, seq2, figure_type='figures')

source

member_frame_movements

 member_frame_movements (movement_series, poses_series, max_clip=3,
                         show=False, condense=True)

For a multi-dancer video, generate individual inter-frame movement series for all of the dancers and plot them.

source

average_frame_movements

 average_frame_movements (movement_series, poses_series, show=False,
                          max_clip=3, video_file=None)

Compute the average and stdv of the inter-frame movement for each frame of a pose sequence with >= 1 dancers. Assumes movement_series is an array of inter-frame movement values, one for each detected pose, with missing poses identified via np.nan, as generated from process_movement_series.

source

plot_interpose_similarity

 plot_interpose_similarity (pose_series, frame_means, frame_stdevs,
                            video_file, show=False, min_clip=0.2)

For multi-pose videos

source

compare_multiple

 compare_multiple (pose_data, method='distance',
                   figure_type='aligned_figures')

For multi-dancer videos: Get the mean and standard deviation of inter-pose similarities for each frame

source

condense_labels

 condense_labels (labels, cluster_map)

Can be used to “collapse” clusters of similar poses into meta-clusters

source

compute_pose_distribution

 compute_pose_distribution (poses_series, labels, descriptors,
                            figure_type='zeroified_figures',
                            cluster_averages=None)

Assign non-clustered poses to clusters (can take a long time) and generate a clustering timeline heatmap of the pose occurrences.

source

render_pose_distribution

 render_pose_distribution (heatmap, poses_series, labels, descriptors,
                           closest_matches=None, show=True,
                           video_file=None, time_index=None,
                           cell_height=120, xlim=None)

Draw a pose cluster timeline based on a precomputed clustering and assigment of non-clustered poses to clusters via compute_pose_distribution() below. Passing in the heatmap from the previous step saves a lot of time.

source

find_nearest_pose

 find_nearest_pose (pose_matrix, cluster_averages)

source

get_cluster_averages_and_indices

 get_cluster_averages_and_indices (labels, descriptors, pose_series,
                                   figure_type='figures', video_file=None,
                                   flip_figures=False)

Average the members of each pose cluster to get the representative “average” pose for the cluster.

source

cluster_poses

 cluster_poses (poses_series, figure_type='aligned_figures',
                min_samples=50)

source

get_feature_vectors

 get_feature_vectors (pose_series, figure_type='aligned_figures',
                      method='distance')

source

average_poses

 average_poses (pose_series, descriptors,
                source_figures='zeroified_figures', flip=True)

source

process_movement_series

 process_movement_series (pose_data, pose_index=-1,
                          figure_type='flipped_figures', video_file=None,
                          method='distance', interpolate=True, viz=True)

Smooth, summarize, visualize movement data for one or more figures across a time series. Also visualize aggregate movement data for each keypoint, if distance matrix method is used.

source

movements_time_series

 movements_time_series (pose_data, pose_index=-1, method='distance',
                        figure_type='flipped_figures', video_file=None)

Calculate a time series of the differences between each pair of poses in a sequence. This works with a single figure (pose_index=0) or all the figures in the video (pose_index=-1). It can be run on its own, but typically this is a helper function for process_movement_series() (below).

source

fill_nans_scipy1

 fill_nans_scipy1 (padata, pkind='linear')

Fill in missing values from a time series, after the first non-NAN value and up to the last non-NAN value. Note that scipy.interpolated.interp1d provides a lot more options (splines, quadratic, etc.)

source

correlate_time_series

 correlate_time_series (pose_data1, pose_data2, method='correlate',
                        figure_type='figures')

source

corr_time_series_matrix

 corr_time_series_matrix (pose_data, method='distance')

Generate a full time-series pose similarity heatmap for all available poses and frames from the video. This code can use either pose characterization approach; in practice, the distance matrix-based analyses take longer to calculate but are more accurate.

source

smooth_series

 smooth_series (x, window_len=11, window='flat')

Smooth a time series via a sliding window average From https://scipy-cookbook.readthedocs.io/items/SignalSmooth.html