dailies_vid.py

# I have a script which runs in the background.
# Every 5 minutes it checks to see if a picture was taken
# for YYYY_MM_DD_HH.jpg. If it does not exist,
# the script takes a photo and checks if faces exist.
# if so it gets saved to YYYY_MM_DD_HH.jpg.

# This is a brief 20-minute excursion into the use of
# perceptual hashes to join 4 months of pictures into
# a video.
# One run is saved at https://youtu.be/9ExS8LDD9BQ

import cv2
import os, sys
from PIL import Image
import random

# python -m pip install --user ImageHash
import imagehash

# Download from https://raw.githubusercontent.com/opencv/opencv/master/data/haarcascades/haarcascade_frontalface_default.xml
FACE_CASCADE_F = '/j/.cache/haarcascade_frontalface_default.xml'

image_folder = '/j/res/profiles/dailies/'
video_name = '/tmp/video.avi'

images = [img for img in os.listdir(image_folder) if img.endswith(".jpg")]
frame = cv2.imread(os.path.join(image_folder, images[0]))
vid_height, vid_width, layers = frame.shape

print("vid_height={}, vid_width={}".format(vid_height, vid_width))

framerate = 8
video = cv2.VideoWriter(video_name, 0, framerate, (vid_width, vid_height))

frames = [
  cv2.imread(os.path.join(image_folder, image)) for image in images
]

# Throw out non-face images
face_cascade = cv2.CascadeClassifier(FACE_CASCADE_F)
def has_face(f):
  global face_cascade
  gray = cv2.cvtColor(f, cv2.COLOR_BGR2GRAY)
  faces = face_cascade.detectMultiScale(gray, 1.1, 5)
  if len(faces) < 1:
    return False
  return True

face_frames = [f for f in frames if has_face(f)]
print("Threw out {} non-face frames...".format(len(frames) - len(face_frames)))
frames = face_frames

for i in range(0, len(frames)):
  height, width, layers = frames[i].shape
  if height != vid_height or width != vid_width:
      frames[i] = cv2.resize(frames[i], (vid_width, vid_height))

print("Read in frames!")

# First sort frames so similar images are next to eachother
# frames = sorted(frames, key=lambda one_frame: hash(imagehash.phash(
#   Image.fromarray( cv2.cvtColor(one_frame, cv2.COLOR_BGR2GRAY) )
# )))

frame_phashes = [
  imagehash.phash( Image.fromarray( img ) )
  for img in frames
]
all_frames_len = len(frames);

some_idx = random.randint(0, len(frames)-1)
print("Started with index {}".format(some_idx))
similar_frames = [ frames[some_idx] ]
added_frame_i = [ some_idx ]

# Skip last 35% of frames b/c they get erratic
max_similar_frames = all_frames_len - int(0.35 * all_frames_len)
while len(similar_frames) < max_similar_frames:
  # Find the most similar frame to similar_frames[-1]...
  if len(similar_frames) % 10 == 0:
    print("{}/{} sorted...".format(len(similar_frames), max_similar_frames))
  
  best_frames_i = 0
  while best_frames_i in added_frame_i:
    best_frames_i += 1

  for i in range(0, len(frames)):
    if i in added_frame_i:
      continue
    # Compute best diff
    diff_best = frame_phashes[added_frame_i[-1]] - frame_phashes[best_frames_i]
    # Compute this diff
    diff_n = frame_phashes[added_frame_i[-1]] - frame_phashes[i]

    if diff_n < diff_best:
      best_frames_i = i

  # Add best
  similar_frames.append( frames[best_frames_i] )
  added_frame_i.append( best_frames_i )

# Now replace frames with similar_frames
frames = similar_frames

print("Sorted!")

# Now join them all
for frame in frames:
    video.write(frame)

cv2.destroyAllWindows()
video.release()

print("Wrote video! ({})".format(video_name))

os.system("mpv '{}'".format(video_name))