Ask HN: Unsure which image extraction techniques to utilise

[zemanel]

For a side project i’m taking a look at how to extract information from Apex Legends (for starters) game play videos. Example screenshot:

https://www.gameprime.org/wp-content/uploads/2019/02/gp-apexlegends-6.jpg

There is a log on top right corner and player status on bottom left corner (health/shields).

Id like to:

- extract <player A username> killed <player B username> with <weapon name> (at time mm:ss) from log, and other information displayed on that log

- from player health extract health/shields percentage (at what time)

Also detect game start/end screens and possibly extract info from them like which characters compose the current player squads.

Apart from OCR, what kind or specific techniques should i learn to implement this ?

[PaulHoule]

If the text is in a predictable position on the screen, this should be easy.

An image processing pipeline like this usually has a filter up front that drops out the environment and leaves a monochrome image of just the text and associated graphics. The text is white and is usually on a background that provides a lot of contrast since it designed to be readable by you!

[zemanel]

Yep makes sense. For the OCR part a colleague suggested something like Tesseract or Google Vision but, what about the shapes, like weapon (type) symbol on top right and health/shield bar percentages and etc ?

Maybe i should just get started small with the OCR part

[PaulHoule]

The weapons symbols are just very big characters.

The health and shield bar percentages could be measured by counting pixels. For instance you could draw a horizontal line across the bar and see how many pixels are lit up.