Re: RGBD SLAM and then RGB Localization

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/RGBD-SLAM-and-then-RGB-Localization-tp489p521.html

The likelihood doesn't tell how many unique correspondences exist between two images. To have many inliers, correspondences should be unique, good and that depth is defined for the features. I don't think the bowEpipolarGeometry is the right thing to use. In your approach, you can have the 3D words of the matched signature directly, "words3" field of the RGBD Signature, that could be used for PnP estimation as the objectPoints argument. The imagePoints could be the matched 2D visual words in the RGB image. Look for example the new PnP code in Memory::computeVisualTransform() on devel branch. Here oldS would be the best match RGB-D image, and newS the RGB image.

cheers