Reading a video directly into Numpy with PyAV (no iterations)

I couldn’t find a solution using PyAV, and uses ffmpeg-python instead.

ffmpeg-python is a Pythonic binding for FFmpeg like PyAV.

The code reads the entire video into a 3D NumPy array of grayscale frames, all at once.

The solution performs the following steps:

• Creates an input video file (for testing).
• Gets the resolution of the video file using "probe".
• Streams the video into byte array.
• Reshape the byte array into n x height x width NumPy array.
• Displays the first frame (for testing).

Here is the code (please read the comments):

import ffmpeg
import numpy as np
from PIL import Image

in_filename = 'in.avi'

"""Build synthetic video, for testing begins:"""
# ffmpeg -y -r 10 -f lavfi -i testsrc=size=160x120:rate=1 -c:v libx264 -t 5 in.mp4
width, height = 160, 120

(
    ffmpeg
    .input('testsrc=size={}x{}:rate=1'.format(width, height), r=10, f='lavfi')
    .output(in_filename, vcodec='libx264', t=5)
    .overwrite_output()
    .run()
)
"""Build synthetic video ends"""

     
# Use FFprobe for getting the resolution of the video frames
p = ffmpeg.probe(in_filename, select_streams='v');
width = p['streams'][0]['width']
height = p['streams'][0]['height']

# https://github.com/kkroening/ffmpeg-python/blob/master/examples/README.md
# Stream the entire video as one large array of bytes
in_bytes, _ = (
    ffmpeg
    .input(in_filename)
    .video # Video only (no audio).
    .output('pipe:', format='rawvideo', pix_fmt='gray')  # Set the output format to raw video in 8 bit grayscale
    .run(capture_stdout=True)
)

n_frames = len(in_bytes) // (height*width)  # Compute the number of frames.
frames = np.frombuffer(in_bytes, np.uint8).reshape(n_frames, height, width) # Reshape buffer to array of n_frames frames (shape of each frame is (height, width)).

im = Image.fromarray(frames[0, :, :])  # Convert first frame to image object
im.show()  # Display the image

---

Output:

---

Update:

Using PyAV:

When using PyAV, we have to decode the video frame by frame.

The main advantage of using PyAV over ffmpeg-python, is that we can use it without the present of FFmpeg CLI (without ffmpeg.exe in Windows).

For reading all video frames into one NumPy array we may use the following stages:

• Decode each frame and convert it to NumPy array.
• Append the NumPy array into a list.
• After iterating all video frames, convert the list to NumPy array that contains all the frames.

---

Code sample (uses OpenCV for showing the frames for testing):

import av
import numpy as np
import cv2

# Build input file using FFmpeg CLI (for testing):
# ffmpeg -y -f lavfi -i testsrc=size=192x108:rate=1:duration=10 -vcodec libx264 -pix_fmt yuv420p myvideo.avi

container = av.open('myvideo.avi')

frames = []  # List of frames - store video frames after converting to NumPy array.

for frame in container.decode(video=0):
    # Decode video frame, and convert to NumPy array in BGR pixel format (use BGR because it used by OpenCV).
    frame = frame.to_ndarray(format='bgr24')  # For Grayscale video, use: frame = frame.to_ndarray(format='gray')
    frames.append(frame)  # Append the frame to the list of frames.

# Convert the list to NumPy array.
# Shape of each frame is (height, width, 3) [for Grayscale the shape is (height, width)]
# the shape of frames is (n_frames, height, width, 3)  [for Grayscale the shape is (n_frames, height, width)]
frames = np.array(frames)

# Show the frames for testing:
for i in range(len(frames)):
    cv2.imshow('frame', frames[i])
    cv2.waitKey(1000)

cv2.destroyAllWindows()