Video Intelligence Output timecodes

[Muhammad Farooq]

Hi All,

As per my observation, video intelligence output is extracting text maximum with the interval of 1/5th of a second (like 0,200,400,600,800 milliseconds). Is there any possibility(parameters to pass etc) so that we can get more precise timecodes (1/40th of a second)? Please help me in this regard.

Best,

[Lei Yang]

Hi Muhammad,

Thanks for using Video Text Detection.

In terms of timestamp unit, it is microsecond.

I am guessing you are talking about number of processed frames per second (processing fps). Our current processing fps is 5fps as you observed. It can handle most of cases with 5fps. (we are still improving it)

May we know what is the motivation to process with high frame rate? To achieve meaningful 1/40 s, input video fps needs to be greater and equal to 40, and computational cost and latency could be 8x compared with now.

Thanks & Regards,

--
© 2018 Google Inc. 1600 Amphitheatre Parkway, Mountain View, CA 94043
 
Email preferences: You received this email because you signed up for the Google Cloud Video Intelligence API Discussion Google Group (cloud-videoi...@googlegroups.com) to participate in discussions with other members of the Google Cloud Video Intelligence API community and the Google Cloud Video Intelligence API Team.
---
You received this message because you are subscribed to the Google Groups "cloud-videointel-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cloud-videointel-d...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cloud-videointel-discuss/47a7432d-f01e-4407-81b8-2e7350bfdb60%40googlegroups.com.

--

Lei Yang (杨蕾) |

Software Engineer, Play |

yanga...@google.com |

650-214-4097

[Muhammad Farooq]

Hi Lei,

Thanks for the quick reply along with the technical clarification.

Actually, we are trying to extract subtitles from the video using google video-ocr service. So, we need these timestamps in a more precise manner like every frame should be processed (25fps). For now with 5fps, Many a time, subs finished earlier than original. Many a time there were gaps created between two subs. I got your point of computational cost vs benefit but if the timestamp is not precise enough, we have to manually correct it and it took a lot more than the computational cost you have mentioned (5x in case of 25fps).

Looking forward for your help if it is possible to automate the manual effort (by processing at 25fps) which could help us a lot from automation point of view.

Thanks again for your comments.

Best,

Farooq

[Lei Yang]

Hi Muhammad,

That is a great use case. 

We will notify you an improvement soon.

Thanks & Regards,

[Muhammad Farooq]

Hi Lei,

Great, thanks! 

Looking forward to hearing from you soon.

Best,

Farooq

--

Senior ML Engineer

Aljazeera Media Network

Doha, Qatar

[Lei Yang]

Hi Muhammad,

Please try the OCR feature, which is improved. And appreciate any feedback.

https://cloud.google.com/video-intelligence/docs/release-notes

Thanks!

[Muhammad Farooq]

Hi Lei,

Thanks, I will check it out and will share the feedback.

Best,

[Muhammad Farooq]

Hi Lei,

I am facing the timeout issue in using the video intelligence text detection service for a video. I am using this service for the past 6 months and usually, it takes on average 6 minutes to process a 25-minute video but now it's not even processing in 30 minutes and timeout occurs. Please note that I am uploading the asset on GCS and then processing. Please help me with this asap.

Best,

Farooq

[Jinjun Qiu]

Hello Farooq,

The latency you observed looks unexpected. Please open an issue at our Public Issue Tracker [1] with your sample video and command for issue reproduction and further investigation.

[1] https://cloud.google.com/support/docs/issue-trackers#issue_reports