Initialization

[anuranbaka]

We need to be able to start the system without initial poses. @hustcalm has been working on this.

[anuranbaka]

I have been doing some investigations on the experimental branch, and have found that the DTAM algo is stable up to scale drift, meaning that aside from the reconstruction scale drifting, the track <--> map loop converges on the right result when given nearby initial poses.

Also the algorithm appears to have two globally stable solutions(neglecting scale drift): one is the correct solution and the other is a reverse depth solution, similar to that seen in the hollow mask illusion or reverspectives. Both solutions happen randomly if random initial displacement and rotation vectors are used between initial frames.

So basically any startup procedure that can distinguish the correct and inverse depth solutions should be sufficient.