Interacting Attention Graph for Single Image Two-Hand Reconstruction

See You Soon: Decoupled Iterative Refinement Framework for Interacting Hands Reconstruction from a Single RGB Image

“…Experiments show that our method significantly outperforms State-Of-The-Art (SOTA) methods (7.52 mm vs 8.79 mm) on the InterHand2.6M dataset [23]. At the same time, we also show the qualitative results of our method on multiple in-the-wild samples [4,39,45,46], from which we can observe that our method has a strong generalization ability.…”

“…The first and second rows show the importance of spatial relationship modeling and mesh-image alignment, respectively. IntagHand [23] is a recently proposed SOTA interacting hands reconstruction method.…”

“…In order to better capture the spatial relationship between hands, Hampali et al [17] extract redundant node features from the visual feature map and use a transformer to perform message passing between nodes of two hands. To achieve pixel-level alignment, Li et al [23] propose a dense attention mechanism to progressively enhance the vertex features of the hand mesh with the image features. However, dense interactions are computationally expensive and suffer from the risk of overfitting.…”

RelightableHands: Efficient Neural Relighting of Articulated Hand Models

“…Facilitated by such parametric models and accurate joint detection methods [52], estimating 3D hand poses is now possible from RGB-D inputs [33,38] or images [34,36,71]. They are also extended to two hands [25,37] and object interactions [16]. While these approaches show impressive robustness, geometric fidelity remains limited.…”

A Survey on Graph Neural Networks and Graph Transformers in Computer Vision: A Task-Oriented Perspective

“…A lightweight GCN-based subnetwork is proposed in [321] for face reconstruction. In addition, hand mesh reconstruction is explored in [322] and [323].…”