Thought we'd share some work-in-progress of distortion improvements for Explorer.
First we acknowledged there are lots of kinds of distortion, such as
"Elevations seem incorrect (white dots don't match road mesh at given location)"
"Dots turn before road"
"Road sidewall not modeled in depth map"
"Dots don't follow almost perfectly straight road. Also occasional misaligned horizon"
"Depth map artefacts (like riding into a wall projected onto it)"
"Well, there are endless depth map issues, some of which we could improve on but many that would still look bad since the photos contain pedestrians and cars that can't be captured in the depth map."
"It bothers me when dots are not following a smooth line and it feels much better when transitions are aligned."
Some of these we were able to improve in our latest release, such as elevations and dots persistently going off-road, by smarter analysis and application of redundant data.
We also experimented with better alignment between photo transitions. This showed promise, but we found a bigger idea that can fix more problems.
The idea is to change the way we render our scenes into something like Pokemon Go in reverse. It will make terrain geometry appear solid between transitions, and be able to incorporate data from multiple image perspectives to avoid the "stretchy" distortion in between images. Here are a few screenshots, with the meshes intentionally highlighted and skymap disabled.
Here's a halfway done video, with solid geometry but still simple image transitions, that shows the benefit on buildings which normally have a bunch of depth errors.
https://drive.google.com/open?id=1-UBfxt...ayCqzMkckx
Note it can't fix the distortion of cars or people or stray trees, which we intentionally captured in this video to show the worst case. But does make it seem more like you're riding through a solid world, especially in VR with stereoscopic depth.
First we acknowledged there are lots of kinds of distortion, such as
"Elevations seem incorrect (white dots don't match road mesh at given location)"
"Dots turn before road"
"Road sidewall not modeled in depth map"
"Dots don't follow almost perfectly straight road. Also occasional misaligned horizon"
"Depth map artefacts (like riding into a wall projected onto it)"
"Well, there are endless depth map issues, some of which we could improve on but many that would still look bad since the photos contain pedestrians and cars that can't be captured in the depth map."
"It bothers me when dots are not following a smooth line and it feels much better when transitions are aligned."
Some of these we were able to improve in our latest release, such as elevations and dots persistently going off-road, by smarter analysis and application of redundant data.
We also experimented with better alignment between photo transitions. This showed promise, but we found a bigger idea that can fix more problems.
The idea is to change the way we render our scenes into something like Pokemon Go in reverse. It will make terrain geometry appear solid between transitions, and be able to incorporate data from multiple image perspectives to avoid the "stretchy" distortion in between images. Here are a few screenshots, with the meshes intentionally highlighted and skymap disabled.
Here's a halfway done video, with solid geometry but still simple image transitions, that shows the benefit on buildings which normally have a bunch of depth errors.
https://drive.google.com/open?id=1-UBfxt...ayCqzMkckx
Note it can't fix the distortion of cars or people or stray trees, which we intentionally captured in this video to show the worst case. But does make it seem more like you're riding through a solid world, especially in VR with stereoscopic depth.