Authors: Jakob Engel, Jürgen Sturm, Daniel Cremers
Year: 2014
- depth of all pixels that have a significant gradient (not all pixels ie. semi dense)
- probabilistic depth map with Gaussian param
- First feature less real time monocular VO
- depth map propagated frame to frame & based on stereo matching of pixels
Stereo matching: we compute the disparity of a pixel (ie its horizontal displacement) which is related to its inverse depth https://en.wikipedia.org/wiki/Computer_stereo_vision
- Uncertainty aware stereo including: photometric error, disparity error and pixel to inverse depth ratio
- depth map regularized by averaging the depth of neighbouring pixels
- dense tracking by minimizing photometric error
Papier trés technique, on s'y perd. L'implémentation a l'air trés galère mais c'est hyper puissant :o Je suis fan