Recherche avancée

Sur d’autres sites (9186)

• h263dec : call get_format() on resolution changes

  8 octobre 2014, par Rémi Denis-Courmont

  h263dec : call get_format() on resolution changes
  Fail safe if the pixel format changes.
  

  • [DBH] libavcodec/h263dec.c

• What is the least CPU-intensive format to pass high resolution frames from ffmpeg to openCV ? [closed]

  3 octobre 2024, par Doctico

  I'm developing an application to process a high-resolution (2560x1440) RTSP stream from an IP camera using OpenCV.

  


  What I've Tried

  


  


  1. OpenCV's VideoCapture :

     


     

     • Performance was poor, even with CAP_PROP_FFMPEG.

     


     


  


  2. FFmpeg with MJPEG :

     


     

     • Decoded the stream as MJPEG and created OpenCV Mats from the image2pipe JPEG buffer.

     


     • Resulted in lower CPU usage for OpenCV but higher for FFmpeg.

     


     


  


  3. Current Approach :

     


     

     • Output raw video in YUV420p format from FFmpeg.

     


     • Construct OpenCV Mats from each frame buffer.

     


     • Achieves low FFmpeg CPU usage and moderately high OpenCV CPU usage.

     


     


  


  


  Current Implementation

  


  import subprocess
import cv2
import numpy as np

def stream_rtsp(rtsp_url):
    # FFmpeg command to stream RTSP and output to pipe
    ffmpeg_command = [
        'ffmpeg',
        '-hwaccel', 'auto',
        '-i', rtsp_url,
        '-pix_fmt', 'yuv420p',  # Use YUV420p format
        '-vcodec', 'rawvideo',
        '-an',  # Disable audio
        '-sn',  # Disable subtitles
        '-f', 'rawvideo',
        '-'  # Output to pipe
    ]

    # Start FFmpeg process
    process = subprocess.Popen(ffmpeg_command, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL)

    # Frame dimensions
    width, height = 2560, 1440
    frame_size = width * height * 3 // 2  # YUV420p uses 1.5 bytes per pixel

    while True:
        # Read raw video frame from FFmpeg output
        raw_frame = process.stdout.read(frame_size)
        if not raw_frame:
            break

        yuv = np.frombuffer(raw_frame, np.uint8).reshape((height * 3 // 2, width))
        frame = cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR_I420)
        
        processFrame(frame)

    # Clean up
    process.terminate()
    cv2.destroyAllWindows()


  


  Question

  


  Are there any other ways to improve performance when processing high-resolution frames from an RTSP stream ?

  


• FFprobe reading incorrect resolution value despite players rendering it correctly

  28 octobre 2024, par Boehmi

  I'm creating a video from a stream with the help of FFMPEG and I also use FFPROBE to gather information for use on a status page like resolution, codecs et cetera.

  



  When FFProbe parses my video for information, I get a resolution value of 544x576 (almost a square !), but an aspect ratio of 16:9.

  



  These values are consistent on both the input stream and my saved video.

  



  When I watch the video in the standard HTML5 Player, VLC or FFPLAY however, I get a video with the proportions of 16:9 and a resolution (measured using an image editing program) of 1024x576 that does look native and not stretched in any way.

  



  Even if I re-encode the video using slightly different codecs, this incorrect resolution value persists, even though every player I use displays it correctly.

  



  This is slightly inconvenient because I am relying on getting the correct resolution value from the video for further processing.

  



  I'm also using a recent FFMPEG+FFPROBE version that was compiled on the 15th of July.

  



  Is this a bug within FFMPEG or is there anything I'm doing wrong ?

  



  Used command lines :

  



  FFMPEG :

  



  ffmpeg -i source -loglevel debug -vcodec copy -acodec copy -bsf:a aac_adtstoasc -movflags +faststart -t 360 -y video.mp4


  



  FFPROBE (I parse the output of this directly and save the values) :

  



  ffprobe -i source -show_format -show_streams 


  



  FFProbe output :

  



  width=544
height=576
coded_width=544
coded_height=576
has_b_frames=2
sample_aspect_ratio=32:17
display_aspect_ratio=16:9


  



  I can see that the sample aspect ratio is different from the display aspect ratio, but how come the video looks proper in 16:9 when it's supposedly encoded at a near square resolution ?