0

I'm using mss, numpy, and OpenCV to make my own recording software, but when I record, all of my videos end up with a 258 bytes file size (no matter the length of the video), and I can't view the video. No frames seem to be in my videos. What am I doing wrong?

Here's my code, any help would be greatly appreciated:

import datetime
import numpy as np
import cv2
import mss
import mss.tools


time_stamp = datetime.datetime.now().strftime('%Y%m%d%H%M%S')
fourcc = cv2.VideoWriter_fourcc("m", "p", "4", "v")
captured_video = cv2.VideoWriter(f"{time_stamp}.mp4", fourcc, 20.0, (1080, 720))

with mss.mss() as sct:
    monitor = {"top": 0, "left": 0, "width": 1080, "height": 720}
    while True:
        img = np.array(sct.grab(monitor))
        img_final0 = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
        img_final = cv2.cvtColor(img_final0, cv2.COLOR_BGR2RGB)
        cv2.imshow('LevelRecorder', img_final)
        capture = captured_video.write(img_final)
        print(capture)
        if cv2.waitKey(10) == ord("q"):
            break
Tiger-222
  • 6,677
  • 3
  • 47
  • 60
RadicalSpace
  • 1
  • 1

2 Answers2

1

The "258 bytes" file size is useful information. Since your files are tiny, VideoWriter actually doesn't even write any video data.

  • Make sure the frames you write() are sized 1080 by 720 and 3-channel (shape (720, 1080, 3)), as you announced in the constructor of the VideoWriter object (argument (1080, 720)).
  • You are missing a captured_video.release() call. MP4 files will be corrupted if they aren't properly finalized, even if video data actually makes it into the file.
  • Those cvtColor operations make no sense. You simply convert back and forth when you don't need to.
  • Look at the Video I/O section in print(cv.getBuildInformation()) and confirm that you have ffmpeg support. Without that, mp4v video and .mp4 container will not be available.

When in doubt, use MJPG fourcc and .avi container. Those two are built into OpenCV, always available.

Christoph Rackwitz
  • 11,317
  • 4
  • 27
  • 36
  • Hmmm... I tried out your first point, and for some reason, my frames were 2160x1440 and had 4 channels, which is not at all what I specified for – RadicalSpace Jul 04 '22 at 20:15
  • I've heard that some screen capturing libraries just give you the entire screen. I'm surprised too that you can tell it a region and it simply disregards that. [docs](https://python-mss.readthedocs.io/examples.html#part-of-the-screen) suggest that this should have worked. check the version of your `mss`. – Christoph Rackwitz Jul 04 '22 at 20:23
  • It actually doesn't seem to be ignoring my inputted values. Now I put 2160x1440 in the monitor line, and the output for img.shape claims that the frames are in 4320x2880. It's taking my values and scaling them up by 2? Also my version of mss is 6.1.0, which is the most recent version – RadicalSpace Jul 05 '22 at 01:35
  • I find it truly peculiar what's going on, I've followed every API documentation or tutorial or post I could find to a tee, yet the result ends up failing for some inexplicable reason. – RadicalSpace Jul 05 '22 at 01:36
  • are you on a mac or something? DPI scaling is often an issue, everywhere. the people who write libraries run on mediocre machines because nobody pays them to have the latest and greatest :P submit a bug to https://github.com/BoboTiG/python-mss/issues but don't expect much. repo's last commit was 6 months ago – Christoph Rackwitz Jul 05 '22 at 07:52
  • *sigh* yes, I am on a mac, I never thought that would be an issue... – RadicalSpace Jul 06 '22 at 00:47
0

At last, I finally found my answer. Turns out, this problem had a super simple solution that I somehow missed this entire time.

As it turns out, either NumPy or Open Cv (I don't know which of the two), really really hates mss, but really likes Python Pillow, so, all you have to do is convert it to an image like Python Pillow, and there you go! Here's my new code (which works for me)

import datetime
from PIL import Image
import numpy as np
import cv2
from mss import mss


time_stamp = datetime.datetime.now().strftime('%Y%m%d%H%M%S')
fourcc = cv2.VideoWriter_fourcc("m", "p", "4", "v")
captured_video = cv2.VideoWriter(f"{time_stamp}.mp4", fourcc, 20.0, (2880, 1800), 3)
sct = mss()

while True:
    img0 = sct.grab(sct.monitors[1])
    img_PIL = Image.frombytes('RGB', img0.size, img0.bgra, 'raw', 'BGRX') #This is where we convert it to a Pillow-like image
    img1= np.array(img_PIL) #Sorry about my variables, I was too lazy to change them :P
    img = cv2.cvtColor(img1, cv2.COLOR_BGR2RGB)
    cv2.imshow('LevelRecorder', img)
    captured_video.write(img)
    if cv2.waitKey(1) == ord("q"):
        cv2.destroyAllWindows()
        break

So there it is! Just use Python Pillow to convert each frame into something NumPy and Open Cv like using Image.frombytes(). Thank you so much Christoph Rackwitz for helping , I can not bring forth into words how much of a great help you have been :)

RadicalSpace
  • 1
  • 1