Use NudeNet to auto generate skp files

[twiclo]

https://github.com/notAI-tech/NudeNet

I'm interested in seeing if I can use the above project to auto generate skp files. Can you point me in the right direction for where I can learn more about the structure of an skp file so I can better generate them programatically?

[fruiz500]

Hello Tanner, I'm glad that you are joining the effort to help folks watch videos safely. The capability of NudeNet fits right in with features newly added to VideoSkip in version 0.5 Skip files are actually text files with a fancy extension. The app itself has a help file that explains how to write the edits into the Edit box, but it does not explain the guts of skip files. There is also a tutorial item on the VideoSkip Exchange (How To section) with pretty much the same content. I am attaching a more or less typical skip file, made for the latest James Bond movie, "No Time to Die". It is edited for sexual content. No attempt was made to edit for violence, alcohol, language, etc., but the process is similar. As of version 0.5, a skip file is made up of four text blocks, with an empty line separating blocks: 1. Screenshot timing: time in hour:minute:second, newline, explanatory text 2. Skips: the format is very similar to that of a .srt subtitle file, except that decimals use a period rather than a comma. 1st line of each skip contains start and end timing in hour:minute:second format, separated by a text-based arrow -->. The parser admits variations on the arrow format. Then a newline, and below a text containing keywords: type of edit (sex, vilolence, etc.), whether it is to be skipped entirely (no extra keyword), blanked, muted, or blurred, a numerical severity value (1, 2 or 3) and for the blur and blank case, an optional set of relative coordinates for a blur spot, as four decimal numbers separated by commas, enclosed in square brackets. The same line can include explanatory text within round brackets, which the parser ignores. Two slashes at the start of the line makes the parser ignore that skip entirely. Keywords, numbers, and bracketed text can appear in any order. After these two lines there is a blank line (optional), and then the next skip following the same format. 3. After all the skips, a JSON-formatted text in curly brackets, where the labels within quotation marks are services the skip file has been synced with, followed by a colon and the numerical value, in seconds, of the time correction needed for that service, relative to the other timings in the file. 4. The screenshot as a data-uri. The image is in jpeg encoding and is 240 pixels tall; width is variable. See if you can get precise location of the offensive content so the coordinates of the blur spot can be determined automatically. Keep up the good work! Best, Paco Ruiz

…

On Thu, Mar 3, 2022 at 10:49 AM Tanner Wilcox ***@***.***> wrote: https://github.com/notAI-tech/NudeNet I'm interested in seeing if I can use the above project to auto generate skp files. Can you point me in the right direction for where I can learn more about the structure of an skp file so I can better generate them programatically? — Reply to this email directly, view it on GitHub <#8>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATF25H4OELA6V7I4VSFHBTU6DUQDANCNFSM5P3AVQ7Q> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Francisco Ruiz Associate Professor MMAE department Illinois Institute of Technology PL24wordLok==qualifications veterans speaker coating physically tapes worldsex folder book mostly distinguished shine enemy sticker focusing arena believe zinc baker thoroughly==PL24wordLok https:*//*www.*youtube.com <http://youtube.com>*/watch?v=BT_tFXpkUR0 get the PassLok privacy app at: https://passlok.com <http://passlok.com>

[speige]

I'm interested in this also.
I think it'd be good to automate the audio analysis for curse words as well.

I have limited experience with AI, but decided to start researching.

Here's my progress so far:
https://github.com/speige/NSFW_Media_Detector

Only does video analysis, not audio yet.
Goes frame-by-frame (slow), could probably be optimized with multi-threading or by skipping frames.
With GPU acceleration, it seems to take ~ 2 hours to process a feature length film.
Does not yet generate the skip files.

I'm unsure how accurate this will be & if it'll still need significant time in manual review, but it seems promising so far.

For the audio analysis, if speech-to-text is difficult (due to background music/etc interference), we could also search for the subtitles online.

The benefit to running analysis is there can be different versions of a video (director's cut, theatrical, unrated, special edition, international language translation, etc) and so syncing a timestamp from online subtitles or someone else's skip file may be difficult.

It's also nice that you could calibrate the analysis to your own preferences as some people may be more sensitive to certain types of content than others.