We propose a novel object-mediated hand interaction system that enables real-time operation with everyday objects on wearable augmented reality (AR) devices. Despite recent advances, both commercial and academic hand interaction techniques remain constrained, typically requiring external hardware or depending exclusively on bare-hand gestures. Motivated by these constraints, we developed an offloading framework that integrates a high-fidelity transformer-based 3D hand reconstruction model with a dynamic gesture recognition network powered by gated recurrent units (GRU). This architecture ensures stable and accurate gesture recognition even during interaction with physical objects. To evaluate its quantitative performance, we collected a custom dataset based on a predefined gesture set, achieving 93.0% accuracy in 5-fold cross-validation. The complete system implemented on Microsoft HoloLens 2 operates at a real-time framerate, and we further analyze the latency of each step in our framework. Through this interaction paradigm, users can experience immersive and intuitive AR in everyday environments with minimal disruption to natural action behavior. Our projects are available at https://github.com/kaist-uvrlab/UnifiedHOInteraction.

We propose Int3DNet, a scene-aware network that predicts 3D intention areas directly from scene geometry and head-hand motion cues, enabling robust human intention prediction without explicit object-level perception. In Mixed Reality (MR), intention prediction is critical as it enables the system to anticipate user actions and respond proactively, reducing interaction delays and ensuring seamless user experiences. Our method employs a cross attention fusion of sparse motion cues and scene point clouds, offering a novel approach that directly interprets the user's spatial intention within the scene. We evaluated Int3DNet on MoGaze and CIRCLE datasets, which are public datasets for full-body human-scene interactions, showing consistent performance across time horizons of up to 1500 ms and outperforming the baselines, even in diverse and unseen scenes. Moreover, we demonstrate the usability of proposed method through a demonstration of efficient visual question answering (VQA) based on intention areas. Int3DNet provides reliable 3D intention areas derived from head-hand motion and scene geometry, thus enabling seamless interaction between humans and MR systems through proactive processing of intention areas.

We introduce SceneLinker, a novel framework that generates compositional 3D scenes via semantic scene graph from RGB sequences. To adaptively experience Mixed Reality (MR) content based on each user's space, it is essential to generate a 3D scene that reflects the real-world layout by compactly capturing the semantic cues of the surroundings. Prior works struggled to fully capture the contextual relationship between objects or mainly focused on synthesizing diverse shapes, making it challenging to generate 3D scenes aligned with object arrangements. We address these challenges by designing a graph network with cross-check feature attention for scene graph prediction and constructing a graph-variational autoencoder (graph-VAE), which consists of a joint shape and layout block for 3D scene generation. Experiments on the 3RScan/3DSSG and SG-FRONT datasets demonstrate that our approach outperforms state-of-the-art methods in both quantitative and qualitative evaluations, even in complex indoor environments and under challenging scene graph constraints. Our work enables users to generate consistent 3D spaces from their physical environments via scene graphs, allowing them to create spatial MR content. Project page is https://scenelinker2026.github.io.

We present a novel method for avatar body generation that cumulatively updates the texture and normal map in real-time. Multiple images or videos have been broadly adopted to create detailed 3D human models that capture more realistic user identities in both Augmented Reality (AR) and Virtual Reality (VR) environments. However, this approach has a higher spatiotemporal cost because it requires a complex camera setup and extensive computational resources. For lightweight reconstruction of personalized avatar bodies, we design a system that progressively captures the texture and normal values using a single RGBD camera to generate the widely-accepted 3D parametric body model, SMPL-X. Quantitatively, our system maintains real-time performance while delivering reconstruction quality comparable to the state-of-the-art method. Moreover, user studies reveal the benefits of real-time avatar creation and its applicability in various collaborative scenarios. By enabling the production of high-fidelity avatars at a lower cost, our method provides more general way to create personalized avatar in AR/VR applications, thereby fostering more expressive self-representation in the metaverse.

We propose a 3D hand tracking system using bare-hand depth inpainting from an RGB-depth image for a hand interacting with an object. The effectiveness of most existing hand-object tracking methods is impeded by the insufficiency of data, which do not include hand data occluded by the object, and their reliance on the information inferred from assuming the specific object type. We generate a sufficiently accurate bare-hand depth image from a hand interacting with an object using a conditional generative adversarial network, which is trained using the synthesized 2D silhouettes of the object to learn the morphology of the hand. We evaluate the proposed approach using a hierarchical particle filter-based hand tracker and prove that our approach utilizing the bare-hand tracker in the hand-object interaction dataset achieve state-of-the-art performance. The generalization of our work will enable visual-tactile interaction that is more natural in various wearable augmented reality applications.