Recognition update
This version improved the algorithm behind recognising signs and translating them into letters by adding a "Recognition Threshold". This means that once a sign was found, it has to pass a specific threshold, which is based on scores gotten after running found landmarks via the Pytorch model. It can be configured in settings and by default set to 80% (which translated to >= 20 recognition score). Setting it above 80% is not reccomended, as the app might not translate the sign at all.
Now once the sign was recognised and it has passed the Recognition Threshold, it is immidiatly added to translated letters. Duplicate letters coming after each other (e.g.: "aabaa") in the translated word are not allowed any more (due to new recognition system).
There have also been many improvements to the camera inself, with slight optimisations (+-10 fps on average -> +-18 fps on average) and fixes to many bugs which appeared on the Release build of the app (not present on Debug build).
Changes
- Fixed bug where on Release build OpenCV wouldn't open
- General optimisations
- FPS counter setting
- Sign recognition and translation algorithm overhaul
- Now there is a Recognition Threshold which has to be passed in order to register a sign as translated (configurable in settings, 80% by default)
- Sign is immidiatly registered and added to translated letters upon passing the threshold (no delay)
- Removed duplicate letter translation
- Color coded hand sceleton based on recognition status (not recognised - red, recognised - green and new recognition - blue)
Installation
- Download the apk file from here or in assets below
- Open and install the apk file
- Done