syanle/ts-commercials-remover

MPEG-TS stream workflow

ts-commercials-remover's Issues

Cutpoints are not precise

Check if ffmpeg commands that copied from losslessCut cut exactly at I-frames
This may help to do a quick but precise cut:
https://video.stackexchange.com/questions/23533/can-you-losslessly-ediit-h264-at-the-gop-level

So the workflow is going to change a little bit.

1. First I need to iterate through all the I-frames, while doing template matching and storing the results. Since the results are from discontinuous frames, I need to refine the junction frames to achieve a precise search. By the way, the matching speed will be dramatically improved for it acting likes as an algorithm called jump search.
2. The denoising process is still necessary, which is a nasty part that I'm most unstatisfied with.
3. The leaving part is tough, as mentioned before, I need to handle a re-encode process to a small part and normal concat process to the remaining part.

TS segs broken when network condition is poor

1. If urlopen reached time-out, the retry decorator won't work, and would lost this seg forever.
2. If urlopen got a resp and it seems all right, any means to check if this seg was not truncated?

urllib2 leads to truncated http response

urllib2 in module m3u8==0.6.0(latest support for Python 2.7) won't add http header "Accept-Encoding: gzip, deflate" automatically, which causes unexpected truncated length of ts segments in some conditions. Notice that the unexpection always return same length 192 (16322 of 24186 bytes). Maybe the length will change when it comes to other encrypted_fileUrl requests. Since Requests support for m3u8 requires Python 3.7, I decide to migrate.

Mis-cut frames

Simply inverting the minor groups is buggy.
If there is a minor gap sandwiched by two mis-recognized groups, the gap will be inverted incorrectly

    _______ads______       _not ads_          ________ads________    
__ ||||||||||||||||||______||_|____|_________|||||||||||||_||||||___

denoised:
__ ||||||||||||||||||________|_______________||||||||||||||||||||___
                             ↑

So, this one will be mis-cut.