We present a real-time method for generating 3D Gaussian head avatars directly from monocular 2D video streams, without per-subject training or 3D morphable model fitting. We train a single model on a large collection of diverse video clips of human heads to generalize across identities, expressions, and poses. Input images are encoded with a pre-trained vision transformer into feature maps, from which both the static and dynamic components are inferred without relying on facial keypoints or parametric models. Dynamic information is injected into the avatar generation process via cross-attention, enabling both self- and cross-identity reenactment. Experiments on the VFHQ dataset show accurate and robust reconstruction performance and improved view point stability compared to the state-of-the-art. We propose a full end-to-end pipeline which is able to create, animate, and render avatars at over 20 fps, enabling new applications in live telepresence and interactive virtual characters.

Current supervised methods for facial landmark detection require a large amount of training data and may suffer from overfitting to specific datasets due to the massive number of parameters. We introduce a semi-supervised method in which the crucial idea is to first generate implicit face knowledge from the large amounts of unlabeled images of faces available today. In a first, completely unsupervised stage, we train an adversarial autoencoder to reconstruct faces via a low-dimensional face embedding. In a second, supervised stage, we interleave the decoder with transfer layers to retask the generation of color images to the prediction of landmark heatmaps. Our framework (3FabRec) achieves state-of-the-art performance on several common benchmarks and, most importantly, is able to maintain impressive accuracy on extremely small training sets down to as few as 10 images. As the interleaved layers only add a low amount of parameters to the decoder, inference runs at several hundred FPS on a GPU.

For any robot, the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Object recognition and categorization is a very challenging problem, as 3-D objects often give rise to ambiguous, 2-D views. Here, we present a perception-driven exploration and recognition scheme for in-hand object recognition implemented on the iCub humanoid robot. In this setup, the robot actively seeks out object views to optimize the exploration sequence. This is achieved by regarding the object recognition problem as a localization problem. We search for the most likely viewpoint position on the viewsphere of all objects. This problem can be solved efficiently using a particle filter that fuses visual cues with associated motor actions. Based on the state of the filter, we can predict the next best viewpoint after each recognition step by searching for the action that leads to the highest expected information gain. We conduct extensive evaluations of the proposed system in simulation as well as on the actual robot and show the benefit of perception-driven exploration over passive, vision-only processes at discriminating between highly similar objects. We demonstrate that objects are recognized faster and at the same time with a higher accuracy.

Interaction with its environment is a key requisite for a humanoid robot. Especially the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Visual object recognition, however, still remains a challenging problem, as three-dimensional objects often give rise to ambiguous, two-dimensional views. Here, we propose a perception-driven, multisensory exploration and recognition scheme to actively resolve ambiguities that emerge at certain viewpoints. We define an efficient method to acquire two-dimensional views in an object-centered task space and sample characteristic views on a view sphere. Information is accumulated during the recognition process and used to select actions expected to be most beneficial in discriminating similar objects. Besides visual information we take into account proprioceptive information to create more reliable hypotheses. Simulation and real-world results clearly demonstrate the efficiency of active, multisensory exploration over passive, visiononly recognition methods.