Forums/Layer Development Platform/Feature Requests
Improvements in layar vision for Model Rotation and Distance
Matti Lahtinen
suggested this on Feb 14 00:38
Greetings, I am doing some experiments into the uses of public AR (research as a media engineering student) using Layar Vision as an available software example.
However from what I have discovered this doesn't work as great as i would expected it to have due to a couple of issues that arise:
- The model, when placed proper, does occasionally does not rotate correctly (it begins to twist around the viewing axis) after the 'angle of attack' to the 'poster' goes beyond (a guesstimate of) 10 degrees from the front of the image (horizontally), requiring resetting the view and occasionally to reload the model (a quick cover of the camera on the phone) to reset the model rotation. Otherwise the vertical rotation is just fine, considering its limit. This could use improvement in tracking the 'poster' or its elements, as the image is fairly well already recognized. *from further experimentation it is reference image dependent, see my next comment for a better solution thats more on the side of the rating system*
- The rendering distance appears make faces and vertices disappear at around 20 meters of range (infact, the distance of the viewer to the 'poster' also effects this, and is not limited to the object in question). This could be improved by making the value twice that. Currently trying to place a large virtual building is difficult, or doing large interiors. I wouldn't mind the constant recorrection however, and the "shakyness" that comes with even larger objects... due to nearly every frame being processed, thought that could use an improvement as well with some form of smoothing. Or is all of this just due to the fact that there is just too
See image for both problems in action in a single example
Currently work around is just to use a smaller model and smaller values or limit it just to a single straight angle and "simulate" distance", however relative rotation is greatly affected. Also using vision instead of regular POI due to the fact that GPS are not as accurate on all phones as what vision has brought on the table.
Using Galaxy S1 - Android 2.3.3 - Layer 6.0.1
Comments latest first
After further testing with a more conventional, actually designed mockup 'reference' yielded slightly different result. Perhaps one of the improvements that could be done is an actual score that should be given by "the ratings assigned by our server based on the quality and clarity" by giving us a more of an detailed indicator (instead of green, yellow and red), on how well the image may fare. As of the moment, two images of the same indicator (yellow) rate as yellow, however when a test on rotation is used, the other fairs much better than the other.
However, the 20 m rendering rule persists, and could be greatly improved on for creating an illusion of depth.