why are affine light parameters so easily to drift?

[vonzy]

I notice that if I remove the constrain for these 2 parameters ( set setting_affineOptModeA & B = 0) in DSO，then these 2 parameters drift severely. And if I use this constrain in my implementation, the solver can't converge to a G-N solution. I wonder why these two parameters are so difficult to fix?

[vonzy]

@NikolausDemmel @JakobEngel

[yuxuanzhou97]

I have the same problem. I try to use the photometric error model applied in DSO, but I find my implementation can't converge at all(through G-N method). While if I just abandon the (a, b) parameters, my result gets much better. I thought I need to set a priority to (a, b), and I add (a=0, b=0) as part of the residuals, giving huge weighting. But it doesn't work as I thought. Maybe I need to learn more about DSO and optimization to handle these two parameters. It's not so easy as I thought. 😢

[vonzy]

I have solved this problem😅

[NikolausDemmel]

I have solved this problem😅

What made it work for you?

[vonzy]

I set huge prior as DSO......