DTWDistances.py

import numpy as np
import fastdtw as fd
import argparse
import csv
import matplotlib.pyplot as plt
from scipy.spatial.distance import euclidean
from scipy.ndimage.filters import gaussian_filter
from scipy.signal import medfilt
from terminalPrintColors import tcolors 

parser = argparse.ArgumentParser(
        description='Dynamic Time Warping Distance Between Pose Sequences')
parser.add_argument('--seq1', default='seq1.csv', help='Sequence one (CSV)')
parser.add_argument('--seq2', default='seq2.csv', help='Sequence two (CSV)')
parser.add_argument('--varthresh', default='0.05', help='Min. variance to use dimension in DTW')
parser.add_argument('--sigma', default='0.5', help='Sigma for Gaussian smoothing')
args = parser.parse_args()

# TODO: check if the given csv does not exist, give meaningful error msg

coordSequences1 = [[] for _ in range(36)] # two dim. list with list for every coordinate
coordSequences2 = [[] for _ in range(36)]
normCoordSequences1 = [[] for _ in range(36)]
normCoordSequences2 = [[] for _ in range(36)]

# read csv file of first sequence to list
with open(args.seq1, 'r') as csvfile:
    reader_seq1 = csv.reader(csvfile, delimiter=',')
    for row in reader_seq1:
        row_f = [float(e) for e in row] # contains strings, cast to float explicitly
        for i in range(36): # 18 key points -> 36 coordinates
            coordSequences1[i].append(row_f[i])

# read csv file of second sequence to list
with open(args.seq2, 'r') as csvfile:
    reader_seq2 = csv.reader(csvfile, delimiter=',')
    for row in reader_seq2:
        row_f = [float(e) for e in row] # contains strings, cast to float explicitly
        for i in range(36): # 18 key points -> 36 coordinates
            coordSequences2[i].append(row_f[i])

compareDimensions = set() # add all dimensions to set that have a variance > threshold in either sequence

if float(args.sigma) > 0.1:
    for i,s in enumerate(coordSequences1):
        s = medfilt(s, 5)
        coordSequences1[i] = s

    for i,s in enumerate(coordSequences2):
        s = medfilt(s, 5)
        coordSequences2[i] = s

for i,kps in enumerate(coordSequences1):
    varSeq = np.var(kps)
    mean = np.mean(kps)
    std = np.std(kps)
    nmkps = [x-mean for x in kps]
    nmstdkps = [x/std for x in nmkps] # TODO: this might divide by zero
    normCoordSequences1[i] = nmstdkps
    print('S1: Variance Signal ', i, ': ' , varSeq)
    if varSeq > float(args.varthresh):
        compareDimensions.add(i)

for i,kps in enumerate(coordSequences2):
    varSeq = np.var(kps)
    mean = np.mean(kps)
    std = np.std(kps)
    nmkps = [x-mean for x in kps]
    nmstdkps = [x/std for x in nmkps] # TODO: this might divide by zero
    normCoordSequences2[i] = nmstdkps
    print('S2: Variance Signal ', i, ': ' , varSeq)
    if varSeq > float(args.varthresh):
        compareDimensions.add(i)

print(compareDimensions)

'''
For some reason, the commumative property is violated if the dimensions
are in a different order... possibly rounding errors. Therefore they have
to be sorted before performing the  DTW.
'''
compareDimensions = sorted(compareDimensions)

f = plt.figure()

plt.subplot(2,1,2)
for i in range(36):
    plt.plot(normCoordSequences1[i], 'xkcd:grey', label=i)

for i in compareDimensions:
    plt.plot(normCoordSequences1[i], 'xkcd:black', label=i, linewidth=4.0)
    plt.plot(normCoordSequences1[i], label=i, linewidth=3.0)

plt.subplot(2,1,1)
for i in range(36):
    plt.plot(normCoordSequences2[i], 'xkcd:grey', label=i)

for i in compareDimensions:
    plt.plot(normCoordSequences2[i], 'xkcd:black', label=i, linewidth=4.0)
    plt.plot(normCoordSequences2[i], label=i, linewidth=3.0)

f.savefig('signals.pdf', bbox_inches='tight')

print(tcolors.HEADER + "USING SELECTED DIMENSIONS " + tcolors.ENDC)
distSum = 0.0

for dim in compareDimensions:
    distance, path = fd.fastdtw(normCoordSequences1[dim], normCoordSequences2[dim], dist=euclidean)
    distSum += distance
    print(distance)
print(tcolors.HEADER + "NORMALIZED DISTANCE: " + tcolors.ENDC + str(distSum))