Re: Monocular mapping on RPi with low depth quality

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/Monocular-mapping-on-RPi-with-low-depth-quality-tp6832p6838.html

Hi,

I'm tried with two different neural networks (fast-depth and mobilepydnet) to compute an approximate 3D image, but in both cases I'm getting very low quality maps

It will be the case, as this is an approximation of the depth image. There are a couple of problems with this setup.
1) Without having ORB-SLAM in the loop to estimate the scale of the resulting depth image from the neural network, the depth image won't have ever the same scale computed by ORB-SLAM, so the pose scale and depth scale won't match.
2) With mono slam, the scale is corrected on loop closures, in this case, the depth image scale won't be corrected when ORB-SLAM will find a loop closure.
3) RTAB-Map doesn't optimize the graph to correct scale-drfit, it is assumed that the real scale is always known in the poses (like with stereo or rgb-d setups). RTAB-Map is not a Monocular SLAM approach.
4) in this case, even if the scale was known, the depth estimation seems not very accurate, there is a lot of distortion.

cheers,
Mathieu