Deep Probabilistic Feature-Metric Tracking

Dense image alignment from RGB-D images remains a critical issue for real-world applications, especially under challenging lighting conditions and in wide baseline setting. In this letter, we propose new framework to learn pixel-wise deep feature map feature-metric uncertainty predicted by Convolutional Neural Network (CNN), which together formulate probabilistic residual of the two-view constraint that can be minimised using Gauss-Newton coarse-to-fine optimisation framework. Furthermore, our network predicts initial pose faster more reliable convergence. The steps are differentiable unrolled train an end-to-end fashion. Due its essence, approach easily couple with other residuals, where show combination ICP. Experimental results demonstrate state-of-the-art performances on TUM dataset 3D rigid object tracking dataset. We further method's robustness convergence qualitatively.