Video-to-video

Algo

The algorithm is composed of 2 steps:

• search for the time shift that gives the PSNR peak:
  • select the frame/field in middle of the reference video sequence
  • search for the frame/field in main which resembles the most to reference frame, i.e. giving the maximum PSNR value
  • calculate the index delta between the reference and main frames/fields and derive delays
• refine the peak value
  • previous PSNR calculation based on one-to-one frame comparison isn't accurate enough and needs to be refined by extending the calculation to video sequences
  • discard 1st and last frames/fields in both streams because they're probably incomplete
  • keep the common portion of content shared by the 2 streams, derived from index delta
  • recalculate the PSNR over these 2 stripped sequences
  • replace the peak value in PSNR graph (hackish)

Results in the UI

Parameter	Explanation
Maximum of PSNR	the maximum of the PSNR between the ref frame/field and every frame/field of the main sequence
Actual delay	the difference between the capture timestamps of the ref frame/field and the most similar main frame/field
Media delay	actual delay converted in media units
RTP delay	the difference between the RTP timestamps of the ref frame/field and the most similar main frame/field

Notes

• positive values means main is later than reference
• ffmpeg returns inf value for perfectly equal images. This is associated to value 100 in graphes
• if video is interlace, then PSNR operation is performed on fields
• the scan type of both main and reference stream must be the same type

Limitations

• decoded frames/fields are .PNG files so pixels are 8-bit per color channel instead of the initial 10-bit