Somewhat Realistic Camera Shake Tracks
======================================

See camerashake.h and camerashake.cpp in libshakez/ subdirectory.
They are the only code you need to use.

The rest of this readme and other scripts are here only for completeness.

Conversion script
-----------------

The conversion script takes in numpy arrays and outputs looped and cross-faded
binary arrays. It basically centers frame poses and then calls

  theta = scipy.linalg.orthogonal_procrustes(reference_pose, frame_pose)

to get the roll vector theta. The script then linearly mixes in the last
second of the recording into the first second, resulting in a crossfade.

To run it:

    pip install -r requirements.txt
    convert_everything.bat

Data
----

I generated the input files in data/ by tracking three points using Blender's
motion tracker feature and saved them to a numpy array with this script:

    import bpy
    import pickle
    import numpy as np

    clip=bpy.data.movieclips['20210308_133421.mp4']
    WIDTH = 1920
    HEIGHT = 1080

    track_points = []

    for ti, track in enumerate(clip.tracking.tracks):
        print(f"[{ti}] Track {track.name}")
        if track.hide:
            print(f"Skipping hidden track {ti}")
            continue
        points = []

        for i, marker in enumerate(track.markers):
            frame = marker.frame
            co = marker.co
            nx, ny = co
            pix = [WIDTH * nx, HEIGHT * (1-ny)] # normalized [0,1] UVs to pixel coordinates
            points.append(pix)
            print(f"[{ti}][{i}] marker @ {frame} is {co}. Pixels {pix}")
        track_points.append(points)


    arr = np.array(track_points)
    np.save(r"fastwalk.npy", arr, allow_pickle=False)

It would've been nice to all this inside Blender but it couldn't export 2D camera poses.

History
-------

2021-07-14 v5: GLSL version with a texture map added. Crossfade fixed and removed DC bias.
2021-07-13 v4: Broken CSV files fixed.
2021-07-12 Hotfix release that correctly computes pose->right vector.
2021-07-12 First release.