How to track your dragon: A Multi-Attentional Framework for real-time RGB-D 6-DOF Object Pose Tracking

We present a novel multi-attentional convolutional architecture to tackle the\nproblem of real-time RGB-D 6D object pose tracking of single, known objects.\nSuch a problem poses multiple challenges originating both from the objects'\nnature and their interaction with their environment, which previous approaches\nhave failed to fully address. The proposed framework encapsulates methods for\nbackground clutter and occlusion handling by integrating multiple parallel soft\nspatial attention modules into a multitask Convolutional Neural Network (CNN)\narchitecture. Moreover, we consider the special geometrical properties of both\nthe object's 3D model and the pose space, and we use a more sophisticated\napproach for data augmentation during training. The provided experimental\nresults confirm the effectiveness of the proposed multi-attentional\narchitecture, as it improves the State-of-the-Art (SoA) tracking performance by\nan average score of 34.03% for translation and 40.01% for rotation, when tested\non the most complete dataset designed, up to date,for the problem of RGB-D\nobject tracking.\n