Recherche avancée

Sur d’autres sites (4712)

• Separate simultaneously changing regions of video into individual videos

  17 juillet 2019, par Elle Fie

  Given a single video stream (up to 4K resolution), where only small displayed portions may change, I’d like to identify these changing sections and create separate video streams, one for each changing section of the input video stream, in real time.

  Note that this is spatial extraction, not time slicing !

  Q1 : Is there a better name to address this process ?

  Q2 : Is this an already solved problem ?

  It seems ImageMagick’s Compare program supports diffing two images, which I can process to identify regions as coordinates for an ffmpeg crop (launched in parallel for each discovered diff region), but this method relies on having a PNG stream to avoid false positive diffs due to lossy encoding. Also, too slow to happen in real time.

  Q3 : Is there any way ffmpeg can dump out the causal regions influencing scene-change detection ?

• What do I have to do to decompress MP4 video with NVDEC directly to a Texture buffer on an NVidia Jetson AGX Xavier computer ?

  13 septembre 2020, par Alexis Wilke

  What I'm trying to do is decompress two MP4 frames (one per NVDEC) and then manipulate them with OpenGL.

  


  Right now, though, this is too slow (definitely not real time), as I have to make copies of 4K images (3840x2160x3 in RGB) and that's just too much data to copy.

  


  I wrote a small benchmark to verify that part. I can only copy between 240 and 250 such buffers per second with memcpy(). That's too slow when the input movies are 60 fps...

  


  I'm able to use the NVDEC chip to decompress to a buffer through ffmpeg, but to place that in a texture, I then have to get the frame out of ffmpeg (copy 1) and then send that image to a texture (copy 2). Do that for two videos, that's 4 copies... 4K is huge ! So the CPUs don't have the time to do that much work 60 times per second.

  


  So right now I'm looking for a way to send the output of the NVDEC directly to a texture. Looking at GStreamer (the gst-launch-1.0), it takes like 3% CPU and can playback a 4K video in real time. What am I doing wrong ?

  


• How to avoid these .bin files to execute ffmpeg library python

  11 avril 2019, par Josemi

  I had this code in Python 3.7 in Windows :

  import numpy as np
  
  import matplotlib.pyplot as plt
  
  from skimage import io, color
  
  import skvideo.io
  
  video = skvideo.io.vread('foreman.avi')

  But the terminal gave me the error : Cannot find installation of real FFmpeg (which comes with ffprobe). Therefore, I followed this guide : Cannot find installation of real FFmpeg (which comes with ffprobe) . However, they say that I had to add the .bin files from /bin directory into my path, and I added them to my workspace and everything worked.

  Now, the question is if there’s a way I don’t have to add the .bin files everytime I want to use the skvideo.io.vread function. I have even paste those files to the ffmpeg directory in C:\ProgramData\Anaconda3\Lib\site-packages\ffmpeg.

  Thank you for your attention.