Despite recent progress in 3D head avatar generation, balancing identity preservation, i.e., reconstruction, with novel poses and expressions, i.e., animation, remains a challenge. Existing methods struggle to adapt Gaussians to varying geometrical deviations across facial regions, resulting in suboptimal quality. To address this, we propose GeoAvatar, a framework for adaptive geometrical Gaussian Splatting. GeoAvatar leverages Adaptive Pre-allocation Stage (APS), an unsupervised method that segments Gaussians into rigid and flexible sets for adaptive offset regularization. Then, based on mouth anatomy and dynamics, we introduce a novel mouth structure and the part-wise deformation strategy to enhance the animation fidelity of the mouth. Finally, we propose a regularization loss for precise rigging between Gaussians and 3DMM faces. Moreover, we release DynamicFace, a video dataset with highly expressive facial motions. Extensive experiments show the superiority of GeoAvatar compared to state-of-the-art methods in reconstruction and novel animation scenarios.

We present a novel learning-based spherical registration method, called SPHARM-Reg, tailored for establishing cortical shape correspondence. SPHARM-Reg aims to reduce warp distortion that can introduce biases in downstream shape analyses. To achieve this, we tackle two critical challenges: (1) joint rigid and non-rigid alignments and (2) rotation-preserving smoothing. Conventional approaches perform rigid alignment only once before a non-rigid alignment. The resulting rotation is potentially sub-optimal, and the subsequent non-rigid alignment may introduce unnecessary distortion. In addition, common velocity encoding schemes on the unit sphere often fail to preserve the rotation component after spatial smoothing of velocity. To address these issues, we propose a diffeomorphic framework that integrates spherical harmonic decomposition of the velocity field with a novel velocity encoding scheme. SPHARM-Reg optimizes harmonic components of the velocity field, enabling joint adjustments for both rigid and non-rigid alignments. Furthermore, the proposed encoding scheme using spherical functions encourages consistent smoothing that preserves the rotation component. In the experiments, we validate SPHARM-Reg on healthy adult datasets. SPHARM-Reg achieves a substantial reduction in warp distortion while maintaining a high level of registration accuracy compared to existing methods. In the clinical analysis, we show that the extent of warp distortion significantly impacts statistical significance.

The identification of cortical sulci is key for understanding functional and structural development of the cortex. While large, consistent sulci (or primary/secondary sulci) receive significant attention in most studies, the exploration of smaller and more variable sulci (or putative tertiary sulci) remains relatively under-investigated. Despite its importance, automatic labeling of cortical sulci is challenging due to (1) the presence of substantial anatomical variability, (2) the relatively small size of the regions of interest (ROIs) compared to unlabeled regions, and (3) the scarcity of annotated labels. In this paper, we propose a novel end-to-end learning framework using a spherical convolutional neural network (CNN). Specifically, the proposed method learns to effectively warp geometric features in a direction that facilitates the labeling of sulci while mitigating the impact of anatomical variability. Moreover, we introduce a guided-attention mechanism that takes into account the extent of deformation induced by the learned warping. This extracts discriminative features that emphasize sulcal ROIs, while suppressing irrelevant information of unlabeled regions. In the experiments, we evaluate the proposed method on 8 sulci of the posterior medial cortex. Our method outperforms existing methods particularly in the putative tertiary sulci. The code is publicly available at https://github.com/Shape-Lab/DSPHARM-Net.

This paper focuses on the transition of automatic speaker verification systems from time delay neural networks (TDNN) to ResNet-based networks. TDNN-based systems use a statistics pooling layer to aggregate temporal information which is suitable for two-dimensional tensors. Even though ResNet-based models produce three-dimensional tensors, they continue to incorporate the statistics pooling layer. However, the reduction in spatial dimensions in ResNet due to convolution operations, including the temporal axis, raises concerns about temporal information loss and its compatibility with statistics pooling. To address this, we introduce Temporal-Bottleneck ResNet (TB-ResNet), a ResNet-based system that can utilize the nature of statistics pooling more effectively by capturing and retaining frame-level contexts through a temporal bottleneck configuration in its building blocks. The performance of TB-ResNets outperforms the original ResNet counterparts on VoxCeleb1, achieving a significant reduction in both the equal error rate and the minimum detection cost function.

This paper presents a new method, called Facial Texture Perceiver. It deals with the task of facial texture recovery from in-the-wild images without 3D supervision. Motivated by their success in various computer vision tasks, we attempt to use transformers for this task. However, capturing high-fidelity facial details requires a large number of mesh vertices and in this case, naively applying vanilla transformer can incur prohibitively high computational and memory costs. We address this challenge by mapping the input with a large number of mesh vertices to a latent space and performing their attention on this space. Also, we introduce input-level inductive biases by injecting the geometry and appearance embeddings as extra inputs. It helps to data-efficiently learn and generalize in-the-wild domains. The resulting architecture enable the application of Transformers to high-resolution facial meshes. Experiments on CelebA, MICC-Florence and MoFA-test datasets demonstrate that our method can accurately reconstruct facial textures, outperforming state-of-the-art methods.

Set reconciliation between different hosts to hold the same dataset is an important prerequisite in numerous distributed applications. Sending the entire dataset to achieve set reconciliation is inefficient if the majority of the data owned by each host is the same. Each host is required to send exclusive elements uniquely included in its set to the other to minimize communication complexity. This paper proposes an efficient algorithm for a host to identify its exclusive elements using the recursive comparison of ternary Bloom filters, each representing the signature of a subset of elements and used for filtering out subsets with identical elements. Hence, subsets with exclusive elements are identified, and elements included in the identified subsets are programmed to an invertible Bloom filter (IBF) to be sent to the other host. Thus, the number of elements programmed in the IBF is significantly reduced. Simulation results show that the proposed algorithm provides excellent performance compared with existing set reconciliation algorithms under the constraint of the same amount of data communication.