Align A Bunch Of Videos. Second file using merge sort for joining the timestamps.

alignment_dtw.py

'''
Align videos using Dynamic Time Warping

Given a set of video files with associated time stamp files,
the program creates a new set of aligned videos.
Copyright (c) <year> <copyright holders>

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

by Daniel Kohlsdorf
'''
import numpy as np
import matplotlib.pyplot as plt
import cv2
import os
import math
import time
import sklearn

class AlignedVideo:
    '''
    Manages aligned read / write of videos
    '''
    def __init__(self):
        self.video = None
        self.writer = None
        self.ts    = []
        self.pos   = 0
        self.last  = None

    def read_video(self, file, output):
        '''
        Open video input and output

        file: the file to read
        output: the output video
        '''
        self.video = cv2.VideoCapture(file)
        codec  = 0x00000021 # TODO Maybe change
        width  = int(self.video.get(3))
        height = int(self.video.get(4))
        fps = int(self.video.get(5)) 
        print(file, output, width, height, fps)
        self.writer = cv2.VideoWriter(output, codec, fps, (width, height))
        self.last = np.zeros((height, width, 3), dtype=np.uint8)

    def read_timestamps(self, file):
        '''
        Read all time stamps from CSV 
        file: path to filename with one timestamp per row        
        '''
        for line in open(file):
            self.ts.append(float(line.strip()))
    
    def current_timestamp(self):
        '''
        Check the current timestamp
        '''
        return self.ts[self.pos]   
    
    def nop(self):
        '''
        Write last image
        '''
        self.writer.write(self.last)

    def next_frame(self):
        '''
        Read next frame from video and write it to the output.
        Also increases pointer to timestamps        
        '''
        _, frame  = self.video.read()
        self.writer.write(frame)
        self.pos += 1
        self.last = frame
        return frame
    
def euclidean(x, y):
    '''
    Squared euclidean distance between two values x and y.
    If the first value is a sequence, we take the minimum distance of y to all elements in x
    x: float of array-like
    y: float
    '''
    if type(x) == list:
        min_dist = float('inf')
        for v in x:
            d = (v - y) ** 2
            if d < min_dist:
                min_dist = d
        return d
    return (x - y) ** 2

def min3(match, insertion, deletion):
    '''
    Calculate the minimum walking direction 
    '''
    if match <= insertion and match <= deletion:
        return (match, -1, -1)
    if insertion < match and insertion < deletion:
        return (insertion, -1, 0)
    if deletion < insertion and deletion < match:
        return (deletion, 0, -1)

def align(x, y, w):
    '''
    Align two sequences using dynamic time warping.
    https://en.wikipedia.org/wiki/Dynamic_time_warping
    
    x: the first sequence, which can also be a collection of sequences.
    y: a 1-dimensional sequence
    w: the warping band see 
    '''
    n = len(x)
    m = len(y)
    w = max(w, abs(n-m + 2)) 

    # Build dynamic programming matrix and save path for back tracking ... 
    dp = np.ones((n + 1, m + 1)) * float('inf')
    dp[0][0] = 0.0
    bp = {}
    for i in range(1, n + 1):
        for j in range(1, m + 1):
            last, di, dj = min3(
                 dp[i - 1][j - 1],
                 dp[i - 1][j],
                 dp[i][j - 1]
            )
            dp[i][j] = euclidean(x[i - 1], y[j - 1]) + last
            bp[(i, j)] = (i + di, j + dj)
    # Back track alignment path and fill gaps with previous sample ...
    aligned_x = []
    aligned_y = []
    i, j = bp[(n, m)]
    aligned_x.append(x[n - 1])
    aligned_y.append(y[m - 1])
    while i > 0 and j > 0:
        aligned_x.append(x[i - 1])
        aligned_y.append(y[j - 1])
        i, j = bp[(i, j)]
    aligned_x.reverse()
    aligned_y.reverse()
    return (aligned_x, aligned_y, dp[n][m])

def merge(sequence, so_far, band):
    '''
    Merge two sequences by appending the aligned samples 
    '''
    (a, b, _) = align(so_far, sequence, band)    
    result = []    
    for k in range(0, len(a)):
        result.append(a[k] + [b[k]])
    return result

def merge_all(sequences):
    '''
    Greedily merge all sequences into one 
    '''
    so_far = [[x] for x in sequences[0]]
    for i in range(1, len(sequences)):
        print("MERGE: {} {}".format(i + 1, len(sequences)))
        band = int(max(len(so_far), len(sequences[i])) / 10)
        so_far = merge(sequences[i], so_far, band)
    return so_far




# Read all video files with their timestamps into annotated video classes
files = {}
path = 'recordings/private/session.017/video/trial1/' # Should be next to the recordings folder ... otherwise adjust path
for file_name in os.listdir(path):
    file = file_name.split('.')[0].replace('_ts', "")    
    if not file in files:
        files[file] = AlignedVideo()
    if 'mkv' in file_name:
        files[file].read_video(path + file_name, "{}_annotation.mp4".format(file))
    else:
        files[file].read_timestamps(path + file_name)

# Merge all time stamps into a single aligned time stamp sequence
l     = len(files)
pos   = {k: i for i, (k, _) in enumerate(files.items())}
names = {i: k for i, (k, _) in enumerate(files.items())}
sequences = [None for i in range(0, len(files))]
for (k, i) in pos.items():
    sequences[i] = files[k].ts
aligned_list = merge_all(sequences)


# In debug mode copy all videos into one image, downsized
DEBUG = True
display = np.zeros((l * 128, 256, 3), dtype=np.uint8)

if DEBUG:
    # plot aligned sequences
    for i in range(0, l):
        plt.plot([aligned_list[j][i] for j in range(0, len(aligned_list))], label=names[i])
    plt.title('Aligned Sequences')
    plt.xlabel('frame')
    plt.ylabel('seconds')
    plt.legend()
    plt.show()

# Align all frames to their respective timestamps
i = 0
n = len(aligned_list)
start = time.time()
for i in range(0, n):    
    for (k, v) in files.items(): 
        if v.current_timestamp() == aligned_list[i][pos[k]]: 
            frame = v.next_frame()
            if DEBUG:
                img = cv2.resize(frame, (256, 128))
                display[pos[k] * 128:(pos[k] + 1) * 128,:, :] = img    
        else:
            v.nop()
    if DEBUG: 
        cv2.imshow(k, display)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break 
    if i % 1000 == 0 and i > 0:
        end  = time.time()
        print("PROCESS: {} / {} in {}[s]".format(i, n, end - start))
        start = time.time()
    i += 1
for (_, v) in files.items():    
    v.writer.release()

alignment_videos.py

import numpy as np
import matplotlib.pyplot as plt
import cv2
import os
import math
import time

class AlignedVideo:
    '''
    Manages aligned read / write of videos
    '''
    def __init__(self):
        self.video = None
        self.writer = None
        self.ts    = []
        self.pos   = 0
        self.last  = None

    def read_video(self, file, output):
        '''
        Open video input and output

        file: the file to read
        output: the output video
        '''
        self.video = cv2.VideoCapture(file)
        codec  = 0x00000021 # TODO Maybe change
        width  = int(self.video.get(3))
        height = int(self.video.get(4))
        fps = int(self.video.get(5))
        print(file, output, width, height, fps)
        self.writer = cv2.VideoWriter(output, codec, fps, (width, height))
        self.last = np.zeros((height, width, 3), dtype=np.uint8)

    def read_timestamps(self, file):
        '''
        Read all time stamps from CSV
        '''
        for line in open(file):
            self.ts.append(float(line.strip()))
    
    def current_timestamp(self):
        '''
        Check the current timestamp
        '''
        return self.ts[self.pos]   
    
    def nop(self):
        '''
        Write last image
        '''
        self.writer.write(self.last)

    def next_frame(self):
        '''
        Read next frame from video and write it to the output.
        Also increases pointer to timestamps        
        '''
        _, frame  = self.video.read()
        self.writer.write(frame)
        self.pos += 1
        self.last = frame
        return frame


# Read all video files with their timestamps into annotated video classes
files = {}
path = 'recordings/private/session.017/video/trial1/' # Should be next to the recordings folder ... otherwise adjust path
for file_name in os.listdir(path):
    file = file_name.split('.')[0].replace('_ts', "")    
    if not file in files:
        files[file] = AlignedVideo()
    if 'mkv' in file_name:
        files[file].read_video(path + file_name, "{}_annotation.mp4".format(file))
    else:
        files[file].read_timestamps(path + file_name)

# Get all possible time stamps sorted as a time stamp master
all_timestamps = set([])
for (k, v) in files.items():
    for t in v.ts:
        all_timestamps.add(t)
all_timestamps = list(all_timestamps)
all_timestamps = sorted(all_timestamps)

# In debug mode copy all videos into one image, downsized
DEBUG = False
l = len(files)
display = np.zeros((l * 128, 256, 3), dtype=np.uint8)
pos = {}
for (i, k) in enumerate(files.keys()):
    pos[k] = i

# Align all frames to their respective timestamps
i = 0
n = len(all_timestamps)
start = time.time()
for t in all_timestamps:
    for (k, v) in files.items():    
        if v.current_timestamp() == t:
            frame = v.next_frame()
            if DEBUG:
                img = cv2.resize(frame, (256, 128))
                display[pos[k] * 128:(pos[k] + 1) * 128,:, :] = img    
        else:
            v.nop()

    if DEBUG: 
        cv2.imshow(k, display)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break 
    if i % 250 == 0 and i > 0:
        end  = time.time()
        print("PROCESS: {} / {} in {}[s]".format(i, n, end - start))
        start = time.time()
    i += 1
for (_, v) in files.items():    
    v.writer.release()

alignment_videos_mergesort.py

import numpy as np
import matplotlib.pyplot as plt
import cv2
import os
import math
import time

class AlignedVideo:
    '''
    Manages aligned read / write of videos
    '''
    def __init__(self):
        self.video = None
        self.writer = None
        self.ts    = []
        self.pos   = 0
        self.last  = None

    def read_video(self, file, output, n_files):
        '''
        Open video input and output
        
        file: the file to read
        output: the output video
        n_files: adjust frame rate by number of files, since we add n - 1 frames per bin (since we almost never match in any video)
        '''
        self.video = cv2.VideoCapture(file)
        codec  = 0x00000021 # TODO Maybe change
        width  = int(self.video.get(3))
        height = int(self.video.get(4))
        fps = int(self.video.get(5)) * (n_files - 1)
        print(file, output, width, height, fps)
        self.writer = cv2.VideoWriter(output, codec, fps, (width, height))
        self.last = np.zeros((height, width, 3), dtype=np.uint8)

    def read_timestamps(self, file):
        '''
        Read all time stamps from CSV
        '''
        for line in open(file):
            self.ts.append(float(line.strip()))
    
    def current_timestamp(self):
        '''
        Check the current timestamp
        '''
        return self.ts[self.pos]   
    
    def nop(self):
        '''
        Write last image
        '''
        self.writer.write(self.last)

    def next_frame(self):
        '''
        Read next frame from video and write it to the output.
        Also increases pointer to timestamps        
        '''
        _, frame  = self.video.read()
        self.writer.write(frame)
        self.pos += 1
        self.last = frame
        return frame


# Read all video files with their timestamps into annotated video classes
files = {}
path = 'recordings/private/session.017/video/trial1/' # Should be next to the recordings folder ... otherwise adjust path
for file_name in os.listdir(path):
    file = file_name.split('.')[0].replace('_ts', "")    
    if not file in files:
        files[file] = AlignedVideo()
    if 'mkv' in file_name:
        files[file].read_video(path + file_name, "{}_annotation.mp4".format(file), 5)
    else:
        files[file].read_timestamps(path + file_name)

def merge(sorted_x, sorted_y):
    result = []
    i = 0
    j = 0
    while i < len(sorted_x) and j < len(sorted_y):
        if sorted_x[i] > sorted_y[j]:
            result.append(sorted_y[j])
            i += 1
        elif sorted_x[i] > sorted_y[j]:
            result.append(sorted_x[i])
            j += 1
        else:
            result.append(sorted_x[i])
            i += 1
            j += 1
    for rest in range(i, len(sorted_x)):
        result.append(sorted_x[rest])
    for rest in range(j, len(sorted_y)):
        result.append(sorted_y[rest])


def merge(sorted_x, sorted_y):
    '''
    Merge two sorted lists into a new sorted list 
    '''
    result = []
    i = 0
    j = 0
    while i < len(sorted_x) and j < len(sorted_y):
        if sorted_x[i] > sorted_y[j]:
            result.append(sorted_y[j])
            j += 1
        elif sorted_x[i] < sorted_y[j]:            
            result.append(sorted_x[i])
            i += 1
        else:
            result.append(sorted_x[i])
            i += 1
            j += 1
    for rest in range(i, len(sorted_x)):
        result.append(sorted_x[rest])
    for rest in range(j, len(sorted_y)):
        result.append(sorted_y[rest])
    return result

# Get all possible time stamps sorted as a time stamp master
keys = list(files.keys())
all_timestamps = files[keys[0]].ts
for i in range(0, len(keys)): 
    all_timestamps = merge(all_timestamps, files[keys[i]].ts)

# In debug mode copy all videos into one image, downsized
DEBUG = True
l = len(files)
display = np.zeros((l * 128, 256, 3), dtype=np.uint8)
pos = {}
for (i, k) in enumerate(files.keys()):
    pos[k] = i

# Align all frames to their respective timestamps
i = 0
n = len(all_timestamps)
start = time.time()
for t in all_timestamps:
    for (k, v) in files.items(): 
        if v.current_timestamp() == t:
            frame = v.next_frame()
            if DEBUG:
                img = cv2.resize(frame, (256, 128))
                display[pos[k] * 128:(pos[k] + 1) * 128,:, :] = img    
        else:
            v.nop()
    if DEBUG: 
        cv2.imshow(k, display)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break 
    if i % 1000 == 0 and i > 0:
        end  = time.time()
        print("PROCESS: {} / {} in {}[s]".format(i, n, end - start))
        start = time.time()
        break
    i += 1
for (_, v) in files.items():    
    v.writer.release()