Pedestrian trajectory prediction is critical for ensuring safety in autonomous driving, surveillance systems, and urban planning applications. While early approaches primarily focus on one-hop pairwise relationships, recent studies attempt to capture high-order interactions by stacking multiple Graph Neural Network (GNN) layers. However, these approaches face a fundamental trade-off: insufficient layers may lead to under-reaching problems that limit the model's receptive field, while excessive depth can result in prohibitive computational costs. We argue that an effective model should be capable of adaptively modeling both explicit one-hop interactions and implicit high-order dependencies, rather than relying solely on architectural depth. To this end, we propose ViTE (Virtual graph Trajectory Expert router), a novel framework for pedestrian trajectory prediction. ViTE consists of two key modules: a Virtual Graph that introduces dynamic virtual nodes to model long-range and high-order interactions without deep GNN stacks, and an Expert Router that adaptively selects interaction experts based on social context using a Mixture-of-Experts design. This combination enables flexible and scalable reasoning across varying interaction patterns. Experiments on three benchmarks (ETH/UCY, NBA, and SDD) demonstrate that our method consistently achieves state-of-the-art performance, validating both its effectiveness and practical efficiency.

We present the first cross-modality distillation framework specifically tailored for single-panoramic-camera Bird’s-Eye-View (BEV) segmentation. Our approach leverages a novel LiDAR image representation fused from range, intensity and ambient channels, together with a voxel-aligned view transformer that preserves spatial fidelity while enabling efficient BEV processing. During training, a high-capacity LiDAR and camera fusion Teacher network extracts both rich spatial and semantic features for cross-modality knowledge distillation into a lightweight Student network that relies solely on a single 360-degree panoramic camera image. Extensive experiments on the Dur360BEV dataset demonstrate that our teacher model significantly outperforms existing camera-based BEV segmentation methods, achieving a 25.6% IoU improvement. Meanwhile, the distilled Student network attains competitive performance with an 8.5% IoU gain and state-of-the-art inference speed of 31.2 FPS. Moreover, evaluations on KITTI-360 (two fisheye cameras) confirm that our distillation framework generalises to diverse camera setups, underscoring its feasibility and robustness. This approach reduces sensor complexity and deployment costs while providing a practical solution for efficient, low-cost BEV segmentation in real-world autonomous driving. The code is available at: https://github.com/Tom-E-Durham/KD360-VoxelBEV.

Action Quality Assessment (AQA) predicts fine-grained execution scores from action videos and is widely applied in sports, rehabilitation, and skill evaluation. Long-term AQA, as in figure skating or rhythmic gymnastics, is especially challenging since it requires modeling extended temporal dynamics while remaining robust to contextual confounders. Existing approaches either depend on costly annotations or rely on unidirectional temporal modeling, making them vulnerable to spurious correlations and unstable long-term representations. To this end, we propose CaFlow, a unified framework that integrates counter-factual de-confounding with bidirectional time-conditioned flow. The Causal Counterfactual Regularization (CCR) module disentangles causal and confounding features in a self-supervised manner and enforces causal robustness through counterfactual interventions, while the BiT-Flow module models forward and backward dynamics with a cycle-consistency constraint to produce smoother and more coherent representations. Extensive experiments on multiple long-term AQA benchmarks demonstrate that CaFlow achieves state-of-the-art performance. Code is available at https://github.com/Harrison21/CaFlow

Abstract Motion tracking has been an important technique for imitating human‐like movement from large‐scale datasets in physics‐based motion synthesis. However, existing approaches focus on tracking either single character or a particular type of interaction, limiting their ability to handle contact‐rich interactions. Extending single‐character tracking approaches suffers from the instability due to the challenge of forces transferred through contacts. Contact‐rich interactions requires levels of control, which places much greater demands on model capacity. To this end, we propose a robust tracking method based on progressive neural network (PNN) where multiple experts are specialized in learning skills of various difficulties. Our method learns to assign training samples to experts automatically without requiring manually scheduling. Both qualitative and quantitative results show that our method delivers more stable motion tracking in densely interactive movements while enabling more efficient model training.

The pervasive issue of bias in AI presents a significant challenge to painting classification, and is getting more serious as these systems become increasingly integrated into tasks like art curation and restoration. Biases, often arising from imbalanced datasets where certain artistic styles dominate, compromise the fairness and accuracy of model predictions, i.e., classifiers are less accurate on rarely seen paintings. While prior research has made strides in improving classification performance, it has largely overlooked the critical need to address these underlying biases, that is, when dealing with out-of-distribution (OOD) data. Our insight highlights the necessity of a more robust approach to bias mitigation in AI models for art classification on biased training data. We propose a novel OOD-informed model bias adaptive sampling method called BOOST (Bias-Oriented OOD Sampling and Tuning). It addresses these challenges by dynamically adjusting temperature scaling and sampling probabilities, thereby promoting a more equitable representation of all classes. We evaluate our proposed approach to the KaoKore and PACS datasets, focusing on the model’s ability to reduce class-wise bias. We further propose a new metric, Same-Dataset OOD Detection Score (SODC), designed to assess class-wise separation and per-class bias reduction. Our method demonstrates the ability to balance high performance with fairness, making it a robust solution for unbiasing AI models in the art domain.

Diffusion models are learning pattern-learning systems to model and sample from data distributions with three functional components namely the forward process, the reverse process, and the sampling process. The components of diffusion models have gained significant attention with many design factors being considered in common practice. Existing reviews have primarily focused on higher-level solutions, covering less on the design fundamentals of components. This study seeks to address this gap by providing a comprehensive and coherent review of seminal designable factors within each functional component of diffusion models. This provides a finer-grained perspective of diffusion models, benefiting future studies in the analysis of individual components, the design factors for different purposes, and the implementation of diffusion models. • A literature review of design fundamentals of diffusion models. • Identified the three functional components of diffusion models. • Examined the key factors and their functionalities of the components. • Discussed the popular designs and seminal works for each factor.

Motion in-betweening is the problem to synthesize movement between keyposes. Traditional research focused primarily on single characters. Extending them to densely interacting characters is highly challenging, as it demands precise spatial-temporal correspondence between the characters to maintain the interaction, while creating natural transitions towards predefined keyposes. In this research, we present a method for long-horizon interaction in-betweening that enables two characters to engage and respond to one another naturally. To effectively represent and synthesize interactions, we propose a novel solution called Cross-Space In-Betweening, which models the interactions of each character across different conditioning representation spaces. We further observe that the significantly increased constraints in interacting characters heavily limit the solution space, leading to degraded motion quality and diminished interaction over time. To enable long-horizon synthesis, we present two solutions to maintain long-term interaction and motion quality, thereby keeping synthesis in the stable region of the solution space. We first sustain interaction quality by identifying periodic interaction patterns through adversarial learning. We further maintain the motion quality by learning to refine the drifted latent space and prevent pose error accumulation. We demonstrate that our approach produces realistic, controllable, and long-horizon in-between motions of two characters with dynamic boxing and dancing actions across multiple keyposes, supported by extensive quantitative evaluations and user studies.

Image inpainting aims to repair a partially damaged image based on the information from known regions of the images. Achieving semantically plausible inpainting results is particularly challenging because it requires the reconstructed regions to exhibit similar patterns to the semanticly consistent regions. This requires a model with a strong capacity to capture long-range dependencies. Existing models struggle in this regard due to the slow growth of receptive field for Convolutional Neural Networks (CNNs) based methods and patch-level interactions in Transformer-based methods, which are ineffective for capturing long-range dependencies. Motivated by this, we propose SEM-Net, a novel visual State Space model (SSM) vision network, modelling corrupted images at the pixel level while capturing long-range dependencies (LRDs) in state space, achieving a linear computational complexity. To address the inherent lack of spatial awareness in SSM, we introduce the Snake Mamba Block (SMB) and Spatially-Enhanced Feed-forward Network. These innovations enable SEM-Net to outperform state-of-the-art inpainting methods on two distinct datasets, showing significant improvements in capturing LRDs and enhancement in spatial consistency. Additionally, SEM-Net achieves state-of-the-art performance on motion deblurring, demonstrating its generalizability. Our source code is available: https://github.com/ChrisChenl023/SEM-Net.

Abstract We propose a real‐time method for reactive motion synthesis based on the known trajectory of input character, predicting instant reactions using only historical, user‐controlled motions. Our method handles the uncertainty of future movements by introducing an intention predictor, which forecasts key joint intentions to make pose prediction more deterministic from the historical interaction. The intention is later encoded into the latent space of its reactive motion, matched with a codebook which represents mappings between input and output. It samples from the categorical distribution for pose generation and strengthens model robustness through adversarial training. Unlike previous offline approaches, the system can recursively generate intentions and reactive motions using feedback from earlier steps, enabling real‐time, long‐term realistic interactive synthesis. Both quantitative and qualitative experiments show our approach outperforms other matching‐based motion synthesis approaches, delivering superior stability and generalisability. In our method, user can also actively influence the outcome by controlling the moving directions, creating a personalised interaction path that deviates from predefined trajectories.

Human-Object Interaction (HOI) recognition in videos requires understanding both visual patterns and geometric relationships as they evolve over time. Visual and geometric features offer complementary strengths. Visual features capture appearance context, while geometric features provide structural patterns. Effectively fusing these multimodal features without compromising their unique characteristics remains challenging. We observe that establishing robust, entity-specific representations before modeling interactions helps preserve the strengths of each modality. Therefore, we hypothesize that a bottom-up approach is crucial for effective multimodal fusion. Following this insight, we propose the Geometric Visual Fusion Graph Neural Network (GeoVis-GNN), which uses dual-attention feature fusion combined with interdependent entity graph learning. It progressively builds from entity-specific representations toward high-level interaction understanding. To advance HOI recognition to real-world scenarios, we introduce the Concurrent Partial Interaction Dataset (MPHOI-120). It captures dynamic multi-person interactions involving concurrent actions and partial engagement. This dataset helps address challenges like complex human-object dynamics and mutual occlusions. Extensive experiments demonstrate the effectiveness of our method across various HOI scenarios. These scenarios include two-person interactions, single-person activities, bimanual manipulations, and complex concurrent partial interactions. Our method achieves state-of-the-art performance.

Pedestrian trajectory prediction aims to forecast future movements based on historical paths. Spatial-temporal (ST) methods often separately model spatial interactions among pedestrians and temporal dependencies of individuals. They overlook the direct impacts of interactions among different pedestrians across various time steps (i.e., high-order cross-time interactions). This limits their ability to capture ST inter-dependencies and hinders prediction performance. To address these limitations, we propose UniEdge with three major designs. Firstly, we introduce a unified ST graph data structure that simplifies high-order cross-time interactions into first-order relationships, enabling the learning of ST inter-dependencies in a single step. This avoids the information loss caused by multi-step aggregation. Secondly, traditional GNNs focus on aggregating pedestrian node features, neglecting the propagation of implicit interaction patterns encoded in edge features. We propose the Edge-to-Edge-Node-to-Node Graph Convolution (E2E-N2N-GCN), a novel dual-graph network that jointly models explicit N2N social interactions among pedestrians and implicit E2E influence propagation across these interaction patterns. Finally, to overcome the limited receptive fields and challenges in capturing long-range dependencies of auto-regressive architectures, we introduce a transformer encoder-based predictor that enables global modeling of temporal correlation. UniEdge outperforms state-of-the-arts on multiple datasets, including ETH, UCY, and SDD.

Long-term Action Quality Assessment (AQA) aims to evaluate the quantitative performance of actions in long videos. However, existing methods face challenges due to domain shifts between the pre-trained large-scale action recognition backbones and the specific AQA task, thereby hindering their performance. This arises since fine-tuning resource-intensive backbones on small AQA datasets is impractical. We address this by identifying two levels of domain shift: task-level, regarding differences in task objectives, and feature-level, regarding differences in important features. For feature-level shifts, which are more detrimental, we propose Progressive Hierarchical Instruction (PHI) with two strategies. First, Gap Minimization Flow (GMF) leverages flow matching to progressively learn a fast flow path that reduces the domain gap between initial and desired features across shallow to deep layers. Additionally, a temporally-enhanced attention module captures long-range dependencies essential for AQA. Second, List-wise Contrastive Regularization (LCR) facilitates coarse-to-fine alignment by comprehensively comparing batch pairs to learn fine-grained cues while mitigating domain shift. Integrating these modules, PHI offers an effective solution. Experiments demonstrate that PHI achieves state-of-the-art performance on three representative long-term AQA datasets, proving its superiority in addressing the domain shift for long-term AQA.

Trajectory prediction allows better decision-making in applications of autonomous vehicles (AVs) or surveillance by predicting the short-term future movement of traffic agents. It is classified into pedestrian or heterogeneous trajectory prediction. The former exploits the relatively consistent behavior of pedestrians, but is limited in real-world scenarios with heterogeneous traffic agents such as cyclists and vehicles. The latter typically relies on extra class label information to distinguish the heterogeneous agents, but such labels are costly to annotate and cannot be generalized to represent different behaviors within the same class of agents. In this work, we introduce the behavioral pseudo-labels that effectively capture the behavior distributions of pedestrians and heterogeneous agents solely based on their motion features, significantly improving the accuracy of trajectory prediction. To implement the framework, we propose the behavioral pseudo-label informed sparse graph convolution network (BP-SGCN) that learns pseudo-labels and informs to a trajectory predictor. For optimization, we propose a cascaded training scheme, in which we first learn the pseudo-labels in an unsupervised manner, and then perform end-to-end fine-tuning on the labels in the direction of increasing the trajectory prediction accuracy. Experiments show that our pseudo-labels effectively model different behavior clusters and improve trajectory prediction. Our proposed BP-SGCN outperforms existing methods using both pedestrian [ETH/UCY, pedestrian-only Stanford Drone Dataset (SDD)] and heterogeneous agent datasets (SDD and Argoverse 1).

The advancement of automated driving technology has led to new challenges in the interaction between automated vehicles and human road users. However, there is currently no complete theory that explains how human road users interact with vehicles, and studying them in real-world settings is often unsafe and time-consuming. This study proposes a 3D Virtual Reality (VR) framework for studying how pedestrians interact with human-driven vehicles. The framework uses VR technology to collect data in a safe and cost-effective way, and deep learning methods are used to predict pedestrian trajectories. Specifically, graph neural networks have been used to model pedestrian future trajectories and the probability of crossing the road. The results of this study show that the proposed framework can be for collecting high-quality data on pedestrian-vehicle interactions in a safe and efficient manner. The data can then be used to develop new theories of human-vehicle interaction and aid the Autonomous Vehicles research.

Contemporary point cloud segmentation approaches largely rely on richly annotated 3D training data. However, it is both time-consuming and challenging to obtain consistently accurate annotations for such 3D scene data. Moreover, there is still a lack of investigation into fully unsupervised scene segmentation for point clouds, especially for holistic 3D scenes. This paper presents U3DS3, as a step towards completely unsupervised point cloud segmentation for any holistic 3D scenes. To achieve this, U3DS3leverages a generalized unsupervised segmentation method for both object and background across both indoor and outdoor static 3D point clouds with no requirement for model pre-training, by leveraging only the inherent information of the point cloud to achieve full 3D scene segmentation. The initial step of our proposed approach involves generating superpoints based on the geometric characteristics of each scene. Subsequently, it undergoes a learning process through a spatial clustering-based methodology, followed by iterative training using pseudo-labels generated in accordance with the cluster centroids. Moreover, by leveraging the invariance and equivariance of the volumetric representations, we apply the geometric transformation on voxelized features to provide two sets of descriptors for robust representation learning. Finally, our evaluation provides state-of-the-art results on the ScanNet and SemanticKITTI, and competitive results on the S3DIS, benchmark datasets.

To improve storage and transmission, images are generally compressed. Vector quantization (VQ) is a popular compression method as it has a high compression ratio that suppresses other compression techniques. Despite this, existing adversarial attack methods on image classification are mostly performed in the pixel domain with few exceptions in the compressed domain, making them less applicable in real-world scenarios. In this paper, we propose a novel one-index attack method in the VQ domain to generate adversarial images by a differential evolution algorithm, successfully resulting in image misclassification in victim models. The one-index attack method modifies a single index in the compressed data stream so that the decompressed image is misclassified. It only needs to modify a single VQ index to realize an attack, which limits the number of perturbed indexes. The proposed method belongs to a semi-black-box attack, which is more in line with the actual attack scenario. We apply our method to attack three popular image classification models, i.e., Resnet, NIN, and VGG16. On average, 55.9% and 77.4% of the images in CIFAR-10 and Fashion MNIST, respectively, are successfully attacked, with a high level of misclassification confidence and a low level of image perturbation.

Existing image inpainting methods leverage convolution-based downsampling approaches to reduce spatial dimensions. This may result in information loss from corrupted images where the available information is inherently sparse, especially for the scenario of large missing regions. Recent advances in self-attention mechanisms within transformers have led to significant improvements in many computer vision tasks including inpainting. However, limited by the computational costs, existing methods cannot fully exploit the efficacy of long-range modelling capabilities of such models. In this paper, we propose an end-to-end High-quality INpainting Transformer, abbreviated as HINT, which consists of a novel mask-aware pixel-shuffle downsampling module (MPD) to preserve the visible information extracted from the corrupted image while maintaining the integrity of the information available for highlevel inferences made within the model. Moreover, we propose a Spatially-activated Channel Attention Layer (SCAL), an efficient self-attention mechanism interpreting spatial awareness to model the corrupted image at multiple scales. To further enhance the effectiveness of SCAL, motivated by recent advanced in speech recognition, we introduce a sandwich structure that places feed-forward networks before and after the SCAL module. We demonstrate the superior performance of HINT compared to contemporary state-of-the-art models on four datasets, CelebA, CelebA-HQ, Places2, and Dunhuang.

In many human-computer interaction applications, fast and accurate hand tracking is necessary for an immersive experience. However, raw hand motion data can be flawed due to issues such as joint occlusions and high-frequency noise, hindering the interaction. Using only current motion for interaction can lead to lag, so predicting future movement is crucial for a faster response. Our solution is the Multi-task Spatial-Temporal Graph Auto-Encoder (Multi-STGAE), a model that accurately denoises and predicts hand motion by exploiting the inter-dependency of both tasks. The model ensures a stable and accurate prediction through denoising while maintaining motion dynamics to avoid over-smoothed motion and alleviate time delays through prediction. A gate mechanism is integrated to prevent negative transfer between tasks and further boost multi-task performance. Multi-STGAE also includes a spatial-temporal graph autoencoder block, which models hand structures and motion coherence through graph convolutional networks, reducing noise while preserving hand physiology. Additionally, we design a novel hand partition strategy and hand bone loss to improve natural hand motion generation. We validate the effectiveness of our proposed method by contributing two large-scale datasets with a data corruption algorithm based on two benchmark datasets. To evaluate the natural characteristics of the denoised and predicted hand motion, we propose two structural metrics. Experimental results show that our method outperforms the state-of-the-art, showcasing how the multi-task framework enables mutual benefits between denoising and prediction.

Whilst the availability of 3D LiDAR point cloud data has significantly grown in recent years, annotation remains expensive and time-consuming, leading to a demand for semi-supervised semantic segmentation methods with application domains such as autonomous driving. Existing work very often employs relatively large segmentation backbone networks to improve segmentation accuracy, at the expense of computational costs. In addition, many use uniform sampling to reduce ground truth data requirements for learning needed, often resulting in sub-optimal performance. To address these issues, we propose a new pipeline that employs a smaller architecture, requiring fewer ground-truth annotations to achieve superior segmentation accuracy compared to contemporary approaches. This is facilitated via a novel Sparse Depthwise Separable Convolution module that significantly reduces the network parameter count while retaining overall task performance. To effectively sub-sample our training data, we propose a new Spatio-Temporal Redundant Frame Downsampling (ST-RFD) method that leverages knowledge of sensor motion within the environment to extract a more diverse subset of training data frame samples. To leverage the use of limited annotated data samples, we further propose a soft pseudo-label method informed by Li-DAR reflectivity. Our method outperforms contemporary semi-supervised work in terms of mIoU, using less labeled data, on the SemanticKITTI (59.5@5%) and ScribbleKITTI (58.1@5%) benchmark datasets, based on a$2.3\times reduction$in model parameters and$64l\times fewer$multiply-add operations whilst also demonstrating significant performance improvement on limited training data (i.e., Less is More).

Generating 3D images of complex objects conditionally from a few 2D views is a difficult synthesis problem, compounded by issues such as domain gap and geometric misalignment. For instance, a unified framework such as Generative Adversarial Networks cannot achieve this unless they explicitly define both a domain-invariant and geometric-invariant joint latent distribution, whereas Neural Radiance Fields are generally unable to handle both issues as they optimize at the pixel level. By contrast, we propose a simple and novel 2D to 3D synthesis approach based on conditional diffusion with vector-quantized codes. Operating in an information-rich code space enables high-resolution 3D synthesis via full-coverage attention across the views. Specifically, we generate the 3D codes (e.g. for CT images) conditional on previously generated 3D codes and the entire codebook of two 2D views (e.g. 2D X-rays). Qualitative and quantitative results demonstrate state-of-the-art performance over specialized methods across varied evaluation criteria, including fidelity metrics such as density, coverage, and distortion metrics for two complex volumetric imagery datasets from in real-world scenarios.

Recently, methods for skeleton-based human activity recognition have been shown to be vulnerable to adversarial attacks. However, these attack methods require either the full knowledge of the victim (i.e. white-box attacks), access to training data (i.e. transfer-based attacks) or frequent model queries (i.e. black-box attacks). All their requirements are highly restrictive, raising the question of how detrimental the vulnerability is. In this paper, we show that the vulnerability indeed exists. To this end, we consider a new attack task: the attacker has no access to the victim model or the training data or labels, where we coin the term hard no-box attack. Specifically, we first learn a motion manifold where we define an adversarial loss to compute a new gradient for the attack, named skeleton-motioninformed (SMI) gradient. Our gradient contains information of the motion dynamics, which is different from existing gradient-based attack methods that compute the loss gradient assuming each dimension in the data is independent. The SMI gradient can augment many gradient-based attack methods, leading to a new family of no-box attack methods. Extensive evaluation and comparison show that our method imposes a real threat to existing classifiers. They also show that the SMI gradient improves the transferability and imperceptibility of adversarial samples in both no-box and transfer-based black-box settings.

Pre-captured immersive environments using omnidirectional cameras provide a wide range of virtual reality applications. Previous research has shown that manipulating the eye height in egocentric virtual environments can significantly affect distance perception and immersion. However, the influence of eye height in pre-captured real environments has received less attention due to the difficulty of altering the perspective after finishing the capture process. To explore this influence, we first propose a pilot study that captures real environments with multiple eye heights and asks participants to judge the egocentric distances and immersion. If a significant influence is confirmed, an effective image-based approach to adapt pre-captured real-world environments to the user’s eye height would be desirable. Motivated by the study, we propose a learning-based approach for synthesizing novel views for omnidirectional images with altered eye heights. This approach employs a multitask architecture that learns depth and semantic segmentation in two formats, and generates high-quality depth and semantic segmentation to facilitate the inpainting stage. With the improved omnidirectional-aware layered depth image, our approach synthesizes natural and realistic visuals for eye height adaptation. Quantitative and qualitative evaluation shows favorable results against state-of-the-art methods, and an extensive user study verifies improved perception and immersion for pre-captured real-world environments.

As human exploration of space continues to progress, the use of Mixed Reality (MR) for simulating microgravity environments and facilitating training in hand-object interaction holds immense practical significance. However, hand-object interaction in microgravity presents distinct challenges compared to terrestrial environments due to the absence of gravity. This results in heightened agility and inherent unpredictability of movements that traditional methods struggle to simulate accurately. To this end, we propose a novel MR-based hand-object interaction system in simulated microgravity environments, leveraging physics-based simulations to enhance the interaction between the user’s real hand and virtual objects. Specifically, we introduce a physics-based hand-object interaction model that combines impulse-based simulation with penetration contact dynamics. This accurately captures the intricacies of hand-object interaction in microgravity. By considering forces and impulses during contact, our model ensures realistic collision responses and enables effective object manipulation in the absence of gravity. The proposed system presents a cost-effective solution for users to simulate object manipulation in microgravity. It also holds promise for training space travelers, equipping them with greater immersion to better adapt to space missions. The system reliability and fidelity test verifies the superior effectiveness of our system compared to the state-of-the-art CLAP system.

Action quality assessment (AQA) automatically evaluates how well humans perform actions in a given video, a technique widely used in fields such as rehabilitation medicine, athletic competitions, and specific skills assessment. However, existing works that uniformly divide the video sequence into small clips of equal length suffer from intra-clip confusion and inter-clip incoherence, hindering the further development of AQA. To address this issue, we propose a hierarchical graph convolutional network (GCN). First, semantic information confusion is corrected through clip refinement, generating the ‘shot’ as the basic action unit. We then construct a scene graph by combining several consecutive shots into meaningful scenes to capture local dynamics. These scenes can be viewed as different procedures of a given action, providing valuable assessment cues. The video-level representation is finally extracted via sequential action aggregation among scenes to regress the predicted score distribution, enhancing discriminative features and improving assessment performance. Experiments on the AQA-7, MTL-AQA, and JIGSAWS datasets demonstrate the superiority of the proposed hierarchical GCN over state-of-the-art methods.

As the most common idiopathic inflammatory myopathy in children, juvenile dermatomyositis (JDM) is characterized by skin rashes and muscle weakness. The childhood myositis assessment scale (CMAS) is commonly used to measure the degree of muscle involvement for diagnosis or rehabilitation monitoring. On the one hand, human diagnosis is not scalable and may be subject to personal bias. On the other hand, automatic action quality assessment (AQA) algorithms cannot guarantee 100% accuracy, making them not suitable for biomedical applications. As a solution, we propose a video-based augmented reality system for human-in-the-loop muscle strength assessment of children with JDM. We first propose an AQA algorithm for muscle strength assessment of JDM using contrastive regression trained by a JDM dataset. Our core insight is to visualize the AQA results as a virtual character facilitated by a 3D animation dataset, so that users can compare the real-world patient and the virtual character to understand and verify the AQA results. To allow effective comparisons, we propose a video-based augmented reality system. Given a feed, we adapt computer vision algorithms for scene understanding, evaluate the optimal way of augmenting the virtual character into the scene, and highlight important parts for effective human verification. The experimental results confirm the effectiveness of our AQA algorithm, and the results of the user study demonstrate that humans can more accurately and quickly assess the muscle strength of children using our system.

Trajectory prediction of road users in real-world scenarios is challenging because their movement patterns are stochastic and complex. Previous pedestrian-oriented works have been successful in modelling the complex interactions among pedestrians, but fail in predicting trajectories when other types of road users are involved (e.g., cars, cyclists, etc.), because they ignore user types. Although a few recent works construct densely connected graphs with user label information, they suffer from superfluous spatial interactions and temporal dependencies. To address these issues, we propose Multiclass-SGCN, a sparse graph convolution network based approach for multi-class trajectory prediction that takes into consideration velocity and agent label information and uses a novel interaction mask to adaptively decide the spatial and temporal connections of agents based on their interaction scores. The proposed approach significantly outperformed state-of-the-art approaches on the Stanford Drone Dataset, providing more realistic and plausible trajectory predictions.

Detecting human-object interactions is essential for comprehensive understanding of visual scenes. In particular, spatial connections between humans and objects are important cues for reasoning interactions. To this end, we propose a skeleton-aware graph convolutional network for human-object interaction detection, named SGCN4HOI. Our network exploits the spatial connections between human keypoints and object keypoints to capture their fine-grained structural interactions via graph convolutions. It fuses such geometric features with visual features and spatial configuration features obtained from human-object pairs. Furthermore, to better preserve the object structural information and facilitate human-object interaction detection, we propose a novel skeleton-based object keypoints representation. The performance of SGCN4HOI is evaluated in the public benchmark V-COCO dataset. Experimental results show that the proposed approach outperforms the state-of-the-art pose-based models and achieves competitive performance against other models.

Single-view depth estimation from omnidirectional images has gained popularity with its wide range of applications such as autonomous driving and scene reconstruction. Although data-driven learning-based methods demonstrate significant potential in this field, scarce training data and ineffective 360 estimation algorithms are still two key limitations hindering accurate estimation across diverse domains. In this work, we first establish a large-scale dataset with varied settings called Depth360 to tackle the training data problem. This is achieved by exploring the use of a plenteous source of data, 360 videos from the internet, using a test-time training method that leverages unique information in each omnidirectional sequence. With novel geometric and temporal constraints, our method generates consistent and convincing depth samples to facilitate single-view estimation. We then propose an end-to-end two-branch multi-task learning network, SegFuse, that mimics the human eye to effectively learn from the dataset and estimate high-quality depth maps from diverse monocular RGB images. With a peripheral branch that uses equirectangular projection for depth estimation and a foveal branch that uses cubemap projection for semantic segmentation, our method predicts consistent global depth while maintaining sharp details at local regions. Experimental results show favorable performance against the state-of-the-art methods.

Acquiring the virtual equivalent of exhibits, such as sculptures, in virtual reality (VR) museums, can be labour-intensive and sometimes infeasible. Deep learning based 3D reconstruction approaches allow us to recover 3D shapes from 2D observations, among which single-view-based approaches can reduce the need for human intervention and specialised equipment in acquiring 3D sculptures for VR museums. However, there exist two challenges when attempting to use the well-researched human reconstruction methods: limited data availability and domain shift. Considering sculptures are usually related to humans, we propose our unsupervised 3D domain adaptation method for adapting a single-view 3D implicit reconstruction model from the source (real-world humans) to the target (sculptures) domain. We have compared the generated shapes with other methods and conducted ablation studies as well as a user study to demonstrate the effectiveness of our adaptation method. We also deploy our results in a VR application.

Existing studies on formation control for unmanned aerial vehicles (UAV) have not considered encircling targets where an optimum coverage of the target is required at all times. Such coverage plays a critical role in many real-world applications such as tracking hostile UAVs. This paper proposes a new path planning approach called the Flux Guided (FG) method, which generates collision-free trajectories for multiple UAVs while maximising the coverage of target(s). Our method enables UAVs to track directly toward a target whilst maintaining maximum coverage. Furthermore, multiple scattered targets can be tracked by scaling the formation during flight. FG is highly scalable since it only requires communication between sub-set of UAVs on the open boundary of the formation’s surface. Experimental results further validate that FG generates UAV trajectories 1.5× shorter than previous work and that trajectory planning for 9 leader/follower UAVs to surround a target in two different scenarios only requires 0.52 s and 0.88 s, respectively. The resulting trajectories are suitable for robotic controls after time-optimal parameterisation; we demonstrate this using a 3d dynamic particle system that tracks the desired trajectories using a PID controller.

Schizophrenia is a severe mental health condition that requires a long and complicated diagnostic process. However, early diagnosis is vital to control symptoms. Deep learning has recently become a popular way to analyse and interpret medical data. Past attempts to use deep learning for schizophrenia diagnosis from brain-imaging data have shown promise but suffer from a large training-application gap — it is difficult to apply lab research to the real world. We propose to reduce this training-application gap by focusing on readily accessible data. We collect a data set of psychiatric observations of patients based on DSM-5 criteria. Because similar data is already recorded in all mental health clinics that diagnose schizophrenia using DSM-5, our method could be easily integrated into current processes as a tool to assist clinicians, whilst abiding by formal diagnostic criteria. To facilitate real-world usage of our system, we show that it is interpretable and robust. Understanding how a machine learning tool reaches its diagnosis is essential to allow clinicians to trust that diagnosis. To interpret the framework, we fuse two complementary attention mechanisms, ‘squeeze and excitation’ and ‘self-attention’, to determine global attribute importance and attribute interactivity, respectively. The model uses these importance scores to make decisions. This allows clinicians to understand how a diagnosis was reached, improving trust in the model. Because machine learning models often struggle to generalise to data from different sources, we perform experiments with augmented test data to evaluate the model’s applicability to the real world. We find that our model is more robust to perturbations, and should therefore perform better in a clinical setting. It achieves 98% accuracy with 10-fold cross-validation.

Abstract 3D car models are heavily used in computer games, visual effects, and even automotive designs. As a result, producing such models with minimal labour costs is increasingly more important. To tackle the challenge, we propose a novel system to reconstruct a 3D car using a single sketch image. The system learns from a synthetic database of 3D car models and their corresponding 2D contour sketches and segmentation masks, allowing effective training with minimal data collection cost. The core of the system is a machine learning pipeline that combines the use of a generative adversarial network (GAN) and lazy learning. GAN, being a deep learning method, is capable of modelling complicated data distributions, enabling the effective modelling of a large variety of cars. Its major weakness is that as a global method, modelling the fine details in the local region is challenging. Lazy learning works well to preserve local features by generating a local subspace with relevant data samples. We demonstrate that the combined use of GAN and lazy learning produces is able to produce high-quality results, in which different types of cars with complicated local features can be generated effectively with a single sketch. Our method outperforms existing ones using other machine learning structures such as the variational autoencoder.

We present DurLAR, a high-fidelity 128-channel 3D LiDAR dataset with panoramic ambient (near infrared) and reflectivity imagery, as well as a sample benchmark task using depth estimation for autonomous driving applications. Our driving platform is equipped with a high resolution 128 channel LiDAR, a 2MPix stereo camera, a lux meter and a GNSSlINS system. Ambient and reflectivity images are made available along with the LiDAR point clouds to facilitate multi-modal use of concurrent ambient and reflectivity scene information. Leveraging DurLAR, with a resolution exceeding that of prior benchmarks, we consider the task of monocular depth estimation and use this increased availability of higher resolution, yet sparse ground truth scene depth information to propose a novel joint supervised/self-supervised loss formulation. We compare performance over both our new DurLAR dataset, the established KITTI benchmark and the Cityscapes dataset. Our evaluation shows our joint use supervised and self-supervised loss terms, enabled via the superior ground truth resolution and availability within DurLAR improves the quantitative and qualitative performance of leading contemporary monocular depth estimation approaches (RMSE = 3.639, SqRel = 0.936).

Hand object interaction in mixed reality (MR) relies on the accurate tracking and estimation of human hands, which provide users with a sense of immersion. However, raw captured hand motion data always contains errors such as joints occlusion, dislocation, high-frequency noise, and involuntary jitter. Denoising and obtaining the hand motion data consistent with the user’s intention are of the utmost importance to enhance the interactive experience in MR. To this end, we propose an end-to-end method for hand motion denoising using the spatial-temporal graph auto-encoder (STGAE). The spatial and temporal patterns are recognized simultaneously by constructing the consecutive hand joint sequence as a spatial-temporal graph. Considering the complexity of the articulated hand structure, a simple yet effective partition strategy is proposed to model the physic-connected and symmetry-connected relationships. Graph convolution is applied to extract structural constraints of the hand, and a self-attention mechanism is to adjust the graph topology dynamically. Combining graph convolution and temporal convolution, a fundamental graph encoder or decoder block is proposed. We finally establish the hourglass residual auto-encoder to learn a manifold projection operation and a corresponding inverse projection through stacking these blocks. In this work, the proposed framework has been successfully used in hand motion data denoising with preserving structural constraints between joints. Extensive quantitative and qualitative experiments show that the proposed method has achieved better performance than the state-of-the-art approaches.

Predicting the movement trajectories of multiple classes of road users in real-world scenarios is a challenging task due to the diverse trajectory patterns. While recent works of pedestrian trajectory prediction successfully modelled the influence of surrounding neighbours based on the relative distances, they are ineffective on multi-class trajectory prediction. This is because they ignore the impact of the implicit correlations between different types of road users on the trajectory to be predicted—for example, a nearby pedestrian has a different level of influence from a nearby car. In this paper, we propose to introduce class information into a graph convolutional neural network to better predict the trajectory of an individual. We embed the class labels of the surrounding objects into the label adjacency matrix (LAM), which is combined with the velocity-based adjacency matrix (VAM) comprised of the objects’ velocity, thereby generating a semantics-guided graph adjacency (SAM). SAM effectively models semantic information with trainable parameters to automatically learn the embedded label features that will contribute to the fixed velocity-based trajectory. Such information of spatial and temporal dependencies is passed to a graph convolutional and temporal convolutional network to estimate the predicted trajectory distributions. We further propose new metrics, known as Average2Displacement Error (aADE) and Average Final Displacement Error (aFDE), that assess network accuracy more accurately. We call our framework Semantics-STGCNN. It consistently shows superior performance to the state-of-the-arts in existing and the newly proposed metrics.

Hand pose estimation, which predicts the spatial location of hand joints, is a fundamental task in VR/AR applications. Although existing methods can recover hand pose competently, the tremor issue occurring in hand motion has not been completely solved. Tremor is an involuntary motion accompanied by a desired gesture or hand motion, leading to hand pose that deviates from user's intentions. Considering the characteristic of tremor motion, we present a novel Graph Neural Network for stable 3D hand pose estimation. The input is depth images. The constraint adjacency matrix is devised in Graph Neural Network for dynamically adjusting the topology of a hand graph during message passing and aggregation. Firstly, since there are rich potential constraints among hand joints, we utilize the constraint adjacency matrix to mine the suitable topology, modeling spatial-temporal constraints of joints and outputting the precise tremor hand pose as the pre-estimation result. Then, for obtaining a stable hand pose, we provide a tremor compensation module based on the constraint adjacency matrix, which exploits the constraint between control points and tremor hand pose. Concretely, the control points represented the voluntary motion are employed as constraints to edit the tremor hand pose. Our extensive quantitative and qualitative experiments show that the proposed method has achieved decent performance for 3D tremor hand pose estimation.

Recurrent neural network (RNN) has become popular for human motion prediction thanks to its ability to capture temporal dependencies. However, it has limited capacity in modeling the complex spatial relationship in the human skeletal structure. In this work, we present a novel diffusion convolutional recurrent predictor for spatial and temporal movement forecasting, with multi-step random walks traversing bidirectionally along an adaptive graph to model interdependency among body joints. In the temporal domain, existing methods rely on a single forward predictor with the produced motion deflecting to the drift route, which leads to error accumulations over time. We propose to supplement the forward predictor with a forward discriminator to alleviate such motion drift in the long term under adversarial training. The solution is further enhanced by a backward predictor and a backward discriminator to effectively reduce the error, such that the system can also look into the past to improve the prediction at early frames. The two-way spatial diffusion convolutions and two-way temporal predictors together form a quadruple network. Furthermore, we train our framework by modeling the velocity from observed motion dynamics instead of static poses to predict future movements that effectively reduces the discontinuity problem at early prediction. Our method outperforms the state of the arts on both 3D and 2D datasets, including the Human3.6M, CMU Motion Capture and Penn Action datasets. The results also show that our method correctly predicts both high-dynamic and low-dynamic moving trends with less motion drift.

In this work, we proposed a novel method, called Local Modified Zernike Moment per unit Mass (LMZMPM), for face recognition, which is invariant to illumination, scaling, noise, in-plane rotation, and translation, along with other orthogonal and inherent properties of the Zernike Moments (ZMs). The proposed LMZMPM is computed for each pixel in a neighborhood of size 3 × 3 , and then considers the complex tuple that contains both the phase and magnitude coefficients of LMZMPM as the extracted features. As it contains both the phase and the magnitude components of the complex feature, it has more information about the image and thus preserves both the edge and structural information. We also propose a hybrid similarity measure, combining the Jaccard Similarity with the L1 distance, and applied to the extracted feature set for classification. The feasibility of the proposed LMZMPM technique on varying illumination has been evaluated on the CMU-PIE and the extended Yale B databases with an average Rank-1 Recognition (R1R) accuracy of 99.8% and 98.66% respectively. To assess the reliability of the method with variations in noise, rotation, scaling, and translation, we evaluate it on the AR database and obtain an average R1R higher than that of recent state-of-the-art methods. The proposed method shows a very high recognition rate on Heterogeneous Face Recognition as well, with 100% on CUFS, and 98.80% on CASIA-HFB.

Data-driven modeling of human motions is ubiquitous in computer graphics and computer vision applications, such as synthesizing realistic motions or recognizing actions. Recent research has shown that such problems can be approached by learning a natural motion manifold using deep learning on a large amount data, to address the shortcomings of traditional data-driven approaches. However, previous deep learning methods can be sub-optimal for two reasons. First, the skeletal information has not been fully utilized for feature extraction. Unlike images, it is difficult to define spatial proximity in skeletal motions in the way that deep networks can be applied for feature extraction. Second, motion is time-series data with strong multi-modal temporal correlations between frames. On the one hand, a frame could be followed by several candidate frames leading to different motions; on the other hand, long-range dependencies exist where a number of frames in the beginning are correlated with a number of frames later. Ineffective temporal modeling would either under-estimate the multi-modality and variance, resulting in featureless mean motion or over-estimate them resulting in jittery motions, which is a major source of visual artifacts. In this paper, we propose a new deep network to tackle these challenges by creating a natural motion manifold that is versatile for many applications. The network has a new spatial component for feature extraction. It is also equipped with a new batch prediction model that predicts a large number of frames at once, such that long-term temporally-based objective functions can be employed to correctly learn the motion multi-modality and variances. With our system, long-duration motions can be predicted/synthesized using an open-loop setup where the motion retains the dynamics accurately. It can also be used for denoising corrupted motions and synthesizing new motions with given control signals. We demonstrate that our system can create superior results comparing to existing work in multiple applications.

This paper addresses the problem of recognizing human-human interaction from skeletal sequences. Existing methods are mainly designed to classify single human action. Many of them simply stack the movement features of two characters to deal with human interaction, while neglecting the abundant relationships between characters. In this paper, we propose a novel two-stream recurrent neural network by adopting the geometric features from both single actions and interactions to describe the spatial correlations with different discriminative abilities. The first stream is constructed under pairwise joint distance (PJD) in a fully-connected mesh to categorize the interactions with explicit distance patterns. To better distinguish similar interactions, in the second stream, we combine PJD with the spatial features from individual joint positions using graph convolutions to detect the implicit correlations among joints, where the joint connections in the graph are adaptive for flexible correlations. After spatial modeling, each stream is fed to a bi-directional LSTM to encode two-way temporal properties. To take advantage of the diverse discriminative power of the two streams, we come up with a late fusion algorithm to combine their output predictions concerning information entropy. Experimental results show that the proposed framework achieves state-of-the-art performance on 3D and comparable performance on 2D interaction datasets. Moreover, the late fusion results demonstrate the effectiveness of improving the recognition accuracy compared with single streams.

Facial makeup style transfer is an extremely challenging sub-field of image-to-image-translation. Due to this difficulty, state-of-the-art results are mostly reliant on the Face Parsing Algorithm, which segments a face into parts in order to easily extract makeup features. However, this algorithm can only work well on high-definition images where facial features can be accurately extracted. Faces in many real-world photos, such as those including a large background or multiple people, are typically of low-resolution, which considerably hinders state-of-the-art algorithms. In this paper, we propose an end-to-end holistic approach to effectively transfer makeup styles between two low-resolution images. The idea is built upon a novel weighted multi-scale spatial attention module, which identifies salient pixel regions on low-resolution images in multiple scales, and uses channel attention to determine the most effective attention map. This design provides two benefits: low-resolution images are usually blurry to different extents, so a multi-scale architecture can select the most effective convolution kernel size to implement spatial attention; makeup is applied on both a macro-level (foundation, fake tan) and a micro-level (eyeliner, lipstick) so different scales can excel in extracting different makeup features. We develop an Augmented CycleGAN network that embeds our attention modules at selected layers to most effectively transfer makeup. Our system is tested with the FBD data set, which consists of many low-resolution facial images, and demonstrate that it outperforms state-of-the-art methods, particularly in transferring makeup for blurry images and partially occluded images.

Abstract By overlaying virtual imagery onto the real world, mixed reality facilitates diverse applications and has drawn increasing attention. Enhancing physical in‐hand objects with a virtual appearance is a key component for many applications that require users to interact with tools such as surgery simulations. However, due to complex hand articulations and severe hand‐object occlusions, resolving occlusions in hand‐object interactions is a challenging topic. Traditional tracking‐based approaches are limited by strong ambiguities from occlusions and changing shapes, while reconstruction‐based methods show a poor capability of handling dynamic scenes. In this article, we propose a novel real‐time optimization system to resolve hand‐object occlusions by spatially reconstructing the scene with estimated hand joints and masks. To acquire accurate results, we propose a joint learning process that shares information between two models and jointly estimates hand poses and semantic segmentation. To facilitate the joint learning system and improve its accuracy under occlusions, we propose an occlusion‐aware RGB‐D hand data set that mitigates the ambiguity through precise annotations and photorealistic appearance. Evaluations show more consistent overlays compared with literature, and a user study verifies a more realistic experience.

In this paper, we propose a deep metric for unifying the representation of mesh saliency detection and non-rigid shape matching. While saliency detection and shape matching are two closely related and fundamental tasks in shape analysis, previous methods approach them separately and independently, failing to exploit their mutually beneficial underlying relationship. In view of the existing gap between saliency and matching, we propose to solve them together using a unified metric representation of surface meshes. We show that saliency and matching can be rigorously derived from our representation as the principal eigenvector and the smoothed Laplacian eigenvectors respectively. Learning the representation jointly allows matching to improve the deformation-invariance of saliency while allowing saliency to improve the feature localization of matching. To parameterize the representation from a mesh, we also propose a deep recurrent neural network (RNN) for effectively integrating multi-scale shape features and a soft-thresholding operator for adaptively enhancing the sparsity of saliency. Results show that by jointly learning from a pair of saliency and matching datasets, matching improves the accuracy of detected salient regions on meshes, which is especially obvious for small-scale saliency datasets, such as those having one to two meshes. At the same time, saliency improves the accuracy of shape matchings among meshes with reduced matching errors on surfaces.

Traditional methods for motion comparison consider features from individual characters. However, the semantic meaning of many human activities is usually defined by the interaction between them, such as a high-five interaction of two characters. There is little success in adapting interaction-based features in activity comparison, as they either do not have a fixed topology or are in high dimensional. In this paper, we propose a unified framework for activity comparison from the interaction point of view. Our new metric evaluates the similarity of interaction by adapting the Earth Mover's Distance onto a customized geometric mesh structure that represents spatial-temporal interactions. This allows us to compare different classes of interactions and discover their intrinsic semantic similarity. We created five interaction databases of different natures, covering both two-characters (synthetic and real-people) and character-object interactions, which are open for public uses. We demonstrate how the proposed metric aligns well with the semantic meaning of the interaction. We also apply the metric in interaction retrieval and show how it outperforms existing ones. The proposed method can be used for unsupervised activity detection in monitoring systems and activity retrieval in smart animation systems.

With the rapid increase in individuals participating in resistance training activities, the number of injuries pertaining to these activities has also grown just as aggressively. Diagnosing the causes of injuries and discomfort requires a large amount of resources from highly experienced physiotherapists. In this paper, we propose a new framework to analyse and visualize movement patterns during performance of four major compound lifts. The analysis generated will be used to efficiently determine whether the exercises are being performed correctly, ensuring anatomy remains within its functional range of motion, in order to prevent strain or discomfort that may lead to injury.

The prior knowledge of real human body shapes and poses is fundamental in computer games and animation (e.g. performance capture). Linear subspaces such as the popular SMPL model have a limited capacity to represent the large geometric variations of human shapes and poses. What is worse is that random sampling from them often produces non-realistic humans because the distribution of real humans is more likely to concentrate on a non-linear manifold instead of the full subspace. Towards this problem, we propose to learn human shape and pose manifolds using a more powerful deep generator network, which is trained to produce samples that cannot be distinguished from real humans by a deep discriminator network. In contrast to previous work that learn both the generator and discriminator in the original geometry spaces, we learn them in the more representative latent spaces discovered by a shape and a pose auto-encoder network respectively. Random sampling from our priors produces higher-quality human shapes and poses. The capacity of our priors is best applied to applications such as virtual human synthesis in games.

Automatic dance synthesis has become more and more popular due to the increasing demand in computer games and animations. Existing research generates dance motions without much consideration for the context of the music. In reality, professional dancers make choreography according to the lyrics and music features. In this research, we focus on a particular genre of dance known as sign dance, which combines gesture-based sign language with full body dance motion. We propose a system to automatically generate sign dance from a piece of music and its corresponding sign gesture. The core of the system is a Sign Dance Model trained by multiple regression analysis to represent the correlations between sign dance and sign gesture/music, as well as a set of objective functions to evaluate the quality of the sign dance. Our system can be applied to music visualization, allowing people with hearing difficulties to understand and enjoy music.

Efficient car shape design is a challenging problem in both the automotive industry and the computer animation/games industry. In this paper, we present a system to reconstruct the 3D car shape from a single 2D sketch image. To learn the correlation between 2D sketches and 3D cars, we propose a Variational Autoencoder deep neural network that takes a 2D sketch and generates a set of multi-view depth & mask images, which are more effective representation comparing to 3D mesh, and can be combined to form the 3D car shape. To ensure the volume and diversity of the training data, we propose a feature-preserving car mesh augmentation pipeline for data augmentation. Since deep learning has limited capacity to reconstruct fine-detail features, we propose a lazy learning approach that constructs a small subspace based on a few relevant car samples in the database. Due to the small size of such a subspace, fine details can be represented effectively with a small number of parameters. With a low-cost optimization process, a high-quality car with detailed features is created. Experimental results show that the system performs consistently to create highly realistic cars of substantially different shape and topology, with a very low computational cost.

We propose a novel end-to-end deep learning framework, capable of 3D human shape reconstruction from a 2D image without the need of a 3D prior parametric model. We employ a “prior-less” representation of the human shape using unordered point clouds. Due to the lack of prior information, comparing the generated and ground truth point clouds to evaluate the reconstruction error is challenging. We solve this problem by proposing an Earth Mover’s Distance (EMD) function to find the optimal mapping between point clouds. Our experimental results show that we are able to obtain a visually accurate estimation of the 3D human shape from a single 2D image, with some inaccuracy for heavily occluded parts.

Makeup face verification in the wild is an important research problem for its popularization in real-world. However, little effort has been made to tackle it in computer vision. In this research, we first build a new database, i.e., Facial Beauty Database (FBD), which contains paired facial images of 8933 subjects without and with makeup in different real-world scenarios. To the best of our knowledge, FBD is the largest makeup face database to date compared with existing databases for facial makeup research. Moreover, we propose a new discriminative marginal metric learning (DMML) algorithm to deal with this problem in the wild. Inspired by the fact that interclass marginal faces are usually more discriminative than interclass nonmarginal faces in learning the discriminative metric space, we use the interclass marginal faces to depict the discriminative information. Simultaneously, we wish that those interclass marginal faces without makeup relations are separated from each other as far as possible, so that more discriminative information between facial images without and with makeup can be exploited for verification. Furthermore, since multiple features could provide comprehensive information in describing the facial representations from diverse points of view and extract more informative cues from facial images, we also introduce a multiview discriminative marginal metric learning (MDMML) algorithm by effectively learning a robust metric space such that multiple features from different points of view can be integrated to effectively enhance the performance of makeup face verification. Experimental results on two real-world makeup face databases are utilized to show the effectiveness of our method and the possibility of verifying the makeup relations from facial images in real-world.

Abstract Emotion is considered to be a core element in performances. In computer animation, both body motions and facial expressions are two popular mediums for a character to express the emotion. However, there has been limited research in studying how to effectively synthesize these two types of character movements using different levels of emotion strength with intuitive control, which is difficult to be modeled effectively. In this work, we explore a common model that can be used to represent the emotion for the applications of body motions and facial expressions synthesis. Unlike previous work that encode emotions into discrete motion style descriptors, we propose a continuous control indicator called emotion strength by controlling which a data‐driven approach is presented to synthesize motions with fine control over emotions. Rather than interpolating motion features to synthesize new motion as in existing work, our method explicitly learns a model mapping low‐level motion features to the emotion strength. Because the motion synthesis model is learned in the training stage, the computation time required for synthesizing motions at run time is very low. We further demonstrate the generality of our proposed framework by editing 2D face images using relative emotion strength. As a result, our method can be applied to interactive applications such as computer games, image editing tools, and virtual reality applications, as well as offline applications such as animation and movie production.

Due to the need to interact with virtual objects, the hand-object interaction has become an important element in mixed reality (MR) applications. In this paper, we propose a novel approach to handle the occlusion of augmented 3D object manipulation with hands by exploiting the nature of hand poses combined with tracking-based and model-based methods, to achieve a complete mixed reality experience without necessities of heavy computations, complex manual segmentation processes or wearing special gloves. The experimental results show a frame rate faster than real-time and a great accuracy of rendered virtual appearances, and a user study verifies a more immersive experience compared to past approaches. We believe that the proposed method can improve a wide range of mixed reality applications that involve hand-object interactions.

Abstract Controlling a crowd using multi‐touch devices appeals to the computer games and animation industries, as such devices provide a high‐dimensional control signal that can effectively define the crowd formation and movement. However, existing works relying on pre‐defined control schemes require the users to learn a scheme that may not be intuitive. We propose a data‐driven gesture‐based crowd control system, in which the control scheme is learned from example gestures provided by different users. In particular, we build a database with pairwise samples of gestures and crowd motions. To effectively generalize the gesture style of different users, such as the use of different numbers of fingers, we propose a set of gesture features for representing a set of hand gesture trajectories. Similarly, to represent crowd motion trajectories of different numbers of characters over time, we propose a set of crowd motion features that are extracted from a Gaussian mixture model. Given a run‐time gesture, our system extracts the K nearest gestures from the database and interpolates the corresponding crowd motions in order to generate the run‐time control. Our system is accurate and efficient, making it suitable for real‐time applications such as real‐time strategy games and interactive animation controls.

Human action recognition is crucial to many practical applications, ranging from human-computer interaction to video surveillance. Most approaches either recognize the human action from a fixed view or require the knowledge of view angle, which is usually not available in practical applications. In this paper, we propose a novel end-to-end framework to jointly learn a view-invariance transfer dictionary and a view-invariant classifier. The result of the process is a dictionary that can project real-world 2D video into a view-invariant sparse representation, as well as a classifier to recognize actions with an arbitrary view. The main feature of our algorithm is the use of synthetic data to extract view-invariance between 3D and 2D videos during the pre-training phase. This guarantees the availability of training data, and removes the hassle of obtaining real-world videos in specific viewing angles. Additionally, for better describing the actions in 3D videos, we introduce a new feature set called the 3D dense trajectories to effectively encode extracted trajectory information on 3D videos. Experimental results on the IXMAS, N-UCLA, i3DPost and UWA3DII datasets show improvements over existing algorithms.

Automatic evaluation of sports skills has been an active research area. However, most of the existing research focuses on low-level features such as movement speed and strength. In this work, we propose a framework for automatic motion analysis and visualization, which allows us to evaluate high-level skills such as the richness of actions, the flexibility of transitions and the unpredictability of action patterns. The core of our framework is the construction and visualization of the posture-based graph that focuses on the standard postures for launching and ending actions, as well as the action-based graph that focuses on the preference of actions and their transition probability. We further propose two numerical indices, the Connectivity Index and the Action Strategy Index, to assess skill level according to the graph. We demonstrate our framework with motions captured from different boxers. Experimental results demonstrate that our system can effectively visualize the strengths and weaknesses of the boxers.

Being marker-free and calibration free, Microsoft Kinect is nowadays widely used in many motion-based applications, such as user training for complex industrial tasks and ergonomics pose evaluation. The major problem of Kinect is the placement requirement to obtain accurate poses, as well as its weakness against occlusions. To improve the robustness of Kinect in interactive motion-based applications, real-time data-driven pose reconstruction has been proposed. The idea is to utilize a database of accurately captured human poses as a prior to optimize the Kinect recognized ones, in order to estimate the true poses performed by the user. The key research problem is to identify the most relevant poses in the database for accurate and efficient reconstruction. In this paper, we propose a new pose reconstruction method based on modelling the pose database with a structure called Filtered Pose Graph, which indicates the intrinsic correspondence between poses. Such a graph not only speeds up the database poses selection process, but also improves the relevance of the selected poses for higher quality reconstruction. We apply the proposed method in a challenging environment of industrial context that involves sub-optimal Kinect placement and a large amount of occlusion. Experimental results show that our real-time system reconstructs Kinect poses more accurately than existing methods.

Various machine learning and data mining tasks in classification require abundant data samples to be labeled for training. Conventional active learning methods aim at labeling the most informative samples for alleviating the labor of the user. Many previous studies in active learning select one sample after another in a greedy manner. However, this is not very effective because the classification models has to be retrained for each newly labeled sample. Moreover, many popular active learning approaches utilize the most uncertain samples by leveraging the classification hyperplane of the classifier, which is not appropriate since the classification hyperplane is inaccurate when the training data are small-sized. The problem of insufficient training data in real-world systems limits the potential applications of these approaches. This paper presents a novel method of active learning called manifold regularized experimental design (MRED), which can label multiple informative samples at one time for training. In addition, MRED gives an explicit geometric explanation for the selected samples to be labeled by the user. Different from existing active learning methods, our method avoids the intrinsic problems caused by insufficiently labeled samples in real-world applications. Various experiments on synthetic datasets, the Yale face database and the Corel image database have been carried out to show how MRED outperforms existing methods.

Fuzzy modelling has been widely and successfully applied to control problems. Traditional fuzzy modelling requires either complete experts' knowledge or large data sets to generate rule bases such that the input spaces can be fully covered. Although fuzzy rule interpolation (FRI) relaxes this requirement by approximating rules using their neighbouring ones, it is still difficult for some real world applications to obtain sufficient experts' knowledge and/or data to generate a reasonable sparse rule base to support FRI. Also, the generated rule bases are usually fixed and therefore cannot support dynamic situations. In order to address these limitations, this paper presents a novel rule base generation and adaptation system to allow the creation of rule bases with minimal a priori knowledge. This is implemented by adding accurate interpolated rules into the rule base guided by a performance index from the feedback mechanism, also considering the rule's previous experience information as a weight factor in the process of rule selection for FRI. In particular, the selection of rules for interpolation in this work is based on a combined metric of the weight factors and the distances between the rules and a given observation, rather than being simply based on the distances. Two digitally simulated scenarios are employed to demonstrate the working of the proposed system, with promising results generated for both rule base generation and adaptation.

Fuzzy inference systems have been successfully applied to many real-world applications. Traditional fuzzy inference systems are only applicable to problems with dense rule bases by which the entire input domain is fully covered, whilst fuzzy rule interpolation (FRI) is also able to work with sparse rule bases that may not cover certain observations. Thanks to its ability to work with fewer rules, fuzzy rule interpolation approaches have also been utilised to reduce system complexity by removing those rules which can be approximated by their neighbouring ones for complex fuzzy models. A number of important fuzzy rule base generation approaches have been proposed in the literature, but the majority of these only target dense rule bases for traditional fuzzy inference systems. This paper proposes a novel sparse fuzzy rule base generation method to support FRI. The approach first identifies important rules that cannot be accurately approximated by their neighbouring ones to initialise the rule base. Then the raw rule base is optimised by fine-tuning the membership functions of the fuzzy sets. Experimentation is conducted to demonstrate the working principles of the proposed system, with results comparable to those of traditional methods.

Human action recognition is an important problem in robotic vision. Traditional recognition algorithms usually require the knowledge of view angle, which is not always available in robotic applications such as active vision. In this paper, we propose a new framework to recognize actions with arbitrary views. A main feature of our algorithm is that view-invariance is learned from synthetic 2D and 3D training data using transfer dictionary learning. This guarantees the availability of training data, and removes the hassle of obtaining real world video in specific viewing angles. The result of the process is a dictionary that can project real world 2D video into a view-invariant sparse representation. This facilitates the training of a view-invariant classifier. Experimental results on the IXMAS and N-UCLA datasets show significant improvements over existing algorithms.

This paper addresses the problem of recognizing human actions captured with depth cameras. Human action recognition is a challenging task as the articulated action data is high dimensional in both spatial and temporal domains. An effective approach to handle this complexity is to divide human body into different body parts according to human skeletal joint positions, and performs recognition based on these part-based feature descriptors. Since different types of features could share some similar hidden structures, and different actions may be well characterized by properties common to all features (sharable structure) and those specific to a feature (specific structure), we propose a joint group sparse regression-based learning method to model each action. Our method can mine the sharable and specific structures among its part-based multiple features meanwhile imposing the importance of these part-based feature structures by joint group sparse regularization, in favor of discriminative part-based feature structure selection. To represent the dynamics and appearance of the human body parts, we employ part-based multiple features extracted from skeleton and depth data respectively. Then, using the group sparse regularization techniques, we have derived an algorithm for mining the key part-based features in the proposed learning framework. The resulting features derived from the learnt weight matrices are more discriminative for multi-task classification. Through extensive experiments on three public datasets, we demonstrate that our approach outperforms existing methods.

Planar shape morphing methods offer solutions to blend two shapes with different silhouettes. A naive method to solve the shape morphing problem is to linearly interpolate the coordinates of each corresponding vertex pair between the source and the target polygons. However, simple linear interpolation sometimes creates intermediate polygons that contain self-intersection, resulting in geometrically incorrect transformations.

Motion analysis and visualization are crucial in sports science for sports training and performance evaluation. While primitive computational methods have been proposed for simple analysis such as postures and movements, few can evaluate the high-level quality of sports players such as their skill levels and strategies. We propose a visualization tool to help visualizing boxers' motions and assess their skill levels. Our system automatically builds a graph-based representation from motion capture data and reduces the dimension of the graph onto a 3D space so that it can be easily visualized and understood. In particular, our system allows easy understanding of the boxer's boxing behaviours, preferred actions, potential strength and weakness. We demonstrate the effectiveness of our system on different boxers' motions. Our system not only serves as a tool for visualization, it also provides intuitive motion analysis that can be further used beyond sports science.

Abstract This paper proposes a new algorithm to produce globally coordinated crowds in an environment with multiple paths and obstacles. Simple greedy crowd control methods easily lead to congestion at bottlenecks within scenes, as the characters do not cooperate with one another. In computer animation, this problem degrades crowd quality especially when ordered behaviour is needed, such as soldiers marching towards a castle. Similarly, in applications such as real‐time strategy games, this often causes player frustration, as the crowd will not move as efficiently as it should. Also, planning of building would usually require visualization of ordered evacuation to maximize the flow. Planning such globally coordinated crowd movement is usually labour intensive. Here, we propose a simple solution that is easy to use and efficient in computation. First, we compute the harmonic field of the environment, taking into account the starting points, goals and obstacles. Based on the field, we represent the topology of the environment using a Reeb Graph, and calculate the maximum capacity for each path in the graph. With the harmonic field and the Reeb Graph, path planning of crowd can be performed using a lightweight algorithm, such that any blocking of one another's paths is minimized. Comparing to previous methods, our system can synthesize globally coordinated crowd with smooth and efficient movement. It also enables control of the crowd with high‐level parameters such as the degree of cooperation and congestion. Finally, the method is scalable to thousands of characters with minimal impact to computation time. It is best applied in interactive crowd synthesis systems such as animation designs and real‐time strategy games.

Smart environments and monitoring systems are popular research areas nowadays due to its potential to enhance the quality of life. Applications such as human behavior analysis and workspace ergonomics monitoring are automated, thereby improving well-being of individuals with minimal running cost. The central problem of smart environments is to understand what the user is doing in order to provide the appropriate support. While it is difficult to obtain information of full body movement in the past, depth camera based motion sensing technology such as Kinect has made it possible to obtain 3D posture without complex setup. This has fused a large number of research projects to apply Kinect in smart environments. The common bottleneck of these researches is the high amount of errors in the detected joint positions, which would result in inaccurate analysis and false alarms. In this paper, we propose a framework that accurately classifies the nature of the 3D postures obtained by Kinect using a max-margin classifier. Different from previous work in the area, we integrate the information about the reliability of the tracked joints in order to enhance the accuracy and robustness of our framework. As a result, apart from general classifying activity of different movement context, our proposed method can classify the subtle differences between correctly performed and incorrectly performed movement in the same context. We demonstrate how our framework can be applied to evaluate the user’s posture and identify the postures that may result in musculoskeletal disorders. Such a system can be used in workplace such as offices and factories to reduce risk of injury. Experimental results have shown that our method consistently outperforms existing algorithms in both activity classification and posture healthiness classification. Due to the low cost and the easy deployment process of depth camera based motion sensors, our framework can be applied widely in home and office to facilitate smart environments.

Content-based image retrieval (CBIR) has attracted much attention during the past decades for its potential practical applications to image database management. A variety of relevance feedback (RF) schemes have been designed to bridge the gap between low-level visual features and high-level semantic concepts for an image retrieval task. In the process of RF, it would be impractical or too expensive to provide explicit class label information for each image. Instead, similar or dissimilar pairwise constraints between two images can be acquired more easily. However, most of the conventional RF approaches can only deal with training images with explicit class label information. In this paper, we propose a novel discriminative semantic subspace analysis (DSSA) method, which can directly learn a semantic subspace from similar and dissimilar pairwise constraints without using any explicit class label information. In particular, DSSA can effectively integrate the local geometry of labeled similar images, the discriminative information between labeled similar and dissimilar images, and the local geometry of labeled and unlabeled images together to learn a reliable subspace. Compared with the popular distance metric analysis approaches, our method can also learn a distance metric but perform more effectively when dealing with high-dimensional images. Extensive experiments on both the synthetic data sets and a real-world image database demonstrate the effectiveness of the proposed scheme in improving the performance of the CBIR.

Depth sensor based 3D human motion estimation hardware such as Kinect has made interactive applications more popular recently. However, it is still challenging to accurately recognize postures from a single depth camera due to the inherently noisy data derived from depth images and self-occluding action performed by the user. In this paper, we propose a new real-time probabilistic framework to enhance the accuracy of live captured postures that belong to one of the action classes in the database. We adopt the Gaussian Process model as a prior to leverage the position data obtained from Kinect and marker-based motion capture system. We also incorporate a temporal consistency term into the optimization framework to constrain the velocity variations between successive frames. To ensure that the reconstructed posture resembles the accurate parts of the observed posture, we embed a set of joint reliability measurements into the optimization framework. A major drawback of Gaussian Process is its cubic learning complexity when dealing with a large database due to the inverse of a covariance matrix. To solve the problem, we propose a new method based on a local mixture of Gaussian Processes, in which Gaussian Processes are defined in local regions of the state space. Due to the significantly decreased sample size in each local Gaussian Process, the learning time is greatly reduced. At the same time, the prediction speed is enhanced as the weighted mean prediction for a given sample is determined by the nearby local models only. Our system also allows incrementally updating a specific local Gaussian Process in real time, which enhances the likelihood of adapting to run-time postures that are different from those in the database. Experimental results demonstrate that our system can generate high quality postures even under severe self-occlusion situations, which is beneficial for real-time applications such as motion-based gaming and sport training.

We propose a new method to compute compatible triangulations of two polygons in order to create a smooth geometric transformation between them. Compared with existing methods, our approach creates triangulations of better quality, that is, triangulations with fewer long thin triangles and Steiner points. This results in visually appealing morphing when transforming the shape from one to another. Our method consists of three stages. First, we compatibly decompose the target and source polygons into a set of sub-polygons, in which each source sub-polygon is triangulated. Second, we map the triangulation of a source sub-polygon onto the corresponding sub-polygon of the target polygon using linear transformation, thereby generating the compatible meshes between the source and the target. Third, we refine the compatible meshes, which can create better quality planar shape morphing with detailed textures. Experimental results show that our method can create compatible meshes of higher quality compared with existing methods, which facilitates smoother morphing process. The proposed algorithm is robust and computationally efficient. It can be applied to produce convincing transformations such as interactive 2D animation creation and special effects in movies.

Due to the recent advancement of computer graphics hardware and software algorithms, deformable characters have become more and more popular in real-time applications such as computer games. While there are mature techniques to generate primary deformation from skeletal movement, simulating realistic and stable secondary deformation such as jiggling of fats remains challenging. On one hand, traditional volumetric approaches such as the finite element method require higher computational cost and are infeasible for limited hardware such as game consoles. On the other hand, while shape matching based simulations can produce plausible deformation in real-time, they suffer from a stiffness problem in which particles either show unrealistic deformation due to high gains, or cannot catch up with the body movement. In this paper, we propose a unified multi-layer lattice model to simulate the primary and secondary deformation of skeleton-driven characters. The core idea is to voxelize the input character mesh into multiple anatomical layers including the bone, muscle, fat and skin. Primary deformation is applied on the bone voxels with lattice-based skinning. The movement of these voxels is propagated to other voxel layers using lattice shape matching simulation, creating a natural secondary deformation. Our multi-layer lattice framework can produce simulation quality comparable to those from other volumetric approaches with a significantly smaller computational cost. It is best to be applied in real-time applications such as console games or interactive animation creation.

Recent work has shown that depth image based 3D posture estimation hardware such as Kinect has made interactive applications more popular. However, it is still challenging to accurately recognize postures from a single depth camera due to the inherently noisy data derived from depth images and self-occluding action performed by the user. While previous research has shown that data-driven methods can be used to reconstruct the correct postures, they usually require a large posture database, which greatly limit the usability for systems with constrained hardware such as game console. To solve this problem, we present a new probabilistic framework to enhance the accuracy of the postures live captured by Kinect. We adopt the Gaussian Process model as a prior to leverage position data obtained from Kinect and marker-based motion capture system. We also incorporate a temporal consistency term into the optimization framework to constrain the velocity variations between successive frames. To ensure that the reconstructed posture resembles the observed input data from Kinect when its tracking result is good, we embed joint reliability into the optimization framework. Experimental results demonstrate that our system can generate high quality postures even under severe self-occlusion situations, which is beneficial for real-time posture based applications such as motion-based gaming and sport training.

ABSTRACT Humans adjust their movements in advance to prepare for the forthcoming action, resulting in efficient and smooth transitions. However, traditional computer animation approaches such as motion graphs simply concatenate a series of actions without taking into account the following one. In this paper, we propose a new method to produce preparation behaviors using reinforcement learning. As an offline process, the system learns the optimal way to approach a target and to prepare for interaction. A scalar value called the level of preparation is introduced, which represents the degree of transition from the initial action to the interacting action. To synthesize the movements of preparation, we propose a customized motion blending scheme based on the level of preparation, which is followed by an optimization framework that adjusts the posture to keep the balance. During runtime, the trained controller drives the character to move to a target with the appropriate level of preparation, resulting in a humanlike behavior. We create scenes in which the character has to move in a complex environment and to interact with objects, such as crawling under and jumping over obstacles while walking. The method is useful not only for computer animation but also for real‐time applications such as computer games, in which the characters need to accomplish a series of tasks in a given environment. Copyright © 2013 John Wiley & Sons, Ltd.

ABSTRACT Human motion variation synthesis is important for crowd simulation and interactive applications to enhance synthesis quality. In this paper, we propose a novel generative probabilistic model to synthesize variations of human motion. Our key idea is to model the conditional distribution of each joint via a multivariate Gaussian process model, namely semiparametric latent factor model (SLFM). SLFM can effectively model the correlations between degrees of freedom (DOFs) of joints rather than dealing with each DOF separately as implemented in existing methods. A detailed evaluation is performed to show that the proposed approach can effectively synthesize variations of different types of motions. Motions generated by our method show a richer variation compared with existing ones. Finally, our user study shows that the synthesized motion has a similar level of naturalness to captured human motions. Our method is best applied in computer games and animations to introduce motion variations. Copyright © 2014 John Wiley & Sons, Ltd.

The degrees of freedom of a crowd is much higher than that provided by a standard user input device. Typically, crowd-control systems require multiple passes to design crowd movements by specifying waypoints, and then defining character trajectories and crowd formation. Such multi-pass control would spoil the responsiveness and excitement of real-time control systems. In this paper, we propose a single-pass algorithm to control a crowd in complex environments. We observe that low-level details in crowd movement are related to interactions between characters and the environment, such as diverging/merging at cross points, or climbing over obstacles. Therefore, we simplify the problem by representing the crowd with a deformable mesh, and allow the user, via multitouch input, to specify high-level movements and formations that are important for context delivery. To help prevent congestion, our system dynamically reassigns characters in the formation by employing a mass transport solver to minimize their overall movement. The solver uses a cost function to evaluate the impact from the environment, including obstacles and areas affecting movement speed. Experimental results show realistic crowd movement created with minimal high-level user inputs. Our algorithm is particularly useful for real-time applications including strategy games and interactive animation creation.

This paper presents a new method to synthesize full body motion for controlling humanoid robots in highly constrained environments. Given a reference motion of the robot and the corresponding environment configuration, the spatial relationships between the robot body parts and the environment objects are extracted as a representation called the Interaction Mesh. Such a representation is then used in adapting the reference motion to an altered environment. By preserving the spatial relationships while satisfying physical constraints, collision-free and well balanced motions can be generated automatically and efficiently. Experimental results show that the proposed method can adapt different full body motions in significantly modified environments. Our method can be applied in precise robotic controls under complicated environments, such as rescue robots in accident scenes and searching robots in highly constrained spaces.

Abstract Creating realistic human movement is a time consuming and labour intensive task. The major difficulty is that the user has to edit individual joints while maintaining an overall realistic and collision free posture. Previous research suggests the use of data‐driven inverse kinematics, such that one can focus on the control of a few joints, while the system automatically composes a natural posture. However, as a common problem of kinematics synthesis, penetration of body parts is difficult to avoid in complex movements. In this paper, we propose a new data‐driven inverse kinematics framework that conserves the topology of the synthesizing postures. Our system monitors and regulates the topology changes using the Gauss Linking Integral (GUI), such that penetration can be efficiently prevented. As a result, complex motions with tight body movements, as well as those involving interaction with external objects, can be simulated with minimal manual intervention. Experimental results show that using our system, the user can create high quality human motion in real‐time by controlling a few joints using a mouse or a multi‐touch screen. The movement generated is both realistic and penetration free. Our system is best applied for interactive motion design in computer animations and games.

The recent advancement of motion recognition using Microsoft Kinect stimulates many new ideas in motion capture and virtual reality applications. Utilizing a pattern recognition algorithm, Kinect can determine the positions of different body parts from the user. However, due to the use of a single-depth camera, recognition accuracy drops significantly when the parts are occluded. This hugely limits the usability of applications that involve interaction with external objects, such as sport training or exercising systems. The problem becomes more critical when Kinect incorrectly perceives body parts. This is because applications have limited information about the recognition correctness, and using those parts to synthesize body postures would result in serious visual artifacts. In this paper, we propose a new method to reconstruct valid movement from incomplete and noisy postures captured by Kinect. We first design a set of measurements that objectively evaluates the degree of reliability on each tracked body part. By incorporating the reliability estimation into a motion database query during run time, we obtain a set of similar postures that are kinematically valid. These postures are used to construct a latent space, which is known as the natural posture space in our system, with local principle component analysis. We finally apply frame-based optimization in the space to synthesize a new posture that closely resembles the true user posture while satisfying kinematic constraints. Experimental results show that our method can significantly improve the quality of the recognized posture under severely occluded environments, such as a person exercising with a basketball or moving in a small room.

With the advancement of motion tracking hardware such as the Microsoft Kinect, synthesizing human-like characters with real-time captured movements becomes increasingly important. Traditional kinematics and dynamics approaches perform sub-optimally when the captured motion is noisy or even incomplete. In this paper, we proposed a unified framework to control physically simulated characters with live captured motion from Kinect. Our framework can synthesize any posture in a physical environment using external forces and torques computed by a PD controller. The major problem of Kinect is the incompleteness of the captured posture, with some degree of freedom (DOF) missing due to occlusions and noises. We propose to search for a best matched posture from a motion database constructed in a dimensionality reduced space, and substitute the missing DOF to the live captured data. Experimental results show that our method can synthesize realistic character movements from noisy captured motion. The proposed algorithm is computationally efficient and can be applied to a wide variety of interactive virtual reality applications such as motion-based gaming, rehabilitation and sport training.

This paper proposes a new methodology for synthesizing animations of multiple characters, allowing them to intelligently compete with one another in dense environments, while still satisfying requirements set by an animator. To achieve these two conflicting objectives simultaneously, our method separately evaluates the competition and collaboration of the interactions, integrating the scores to select an action that maximizes both criteria. We extend the idea of min-max search, normally used for strategic games such as chess. Using our method, animators can efficiently produce scenes of dense character interactions such as those in collective sports or martial arts. The method is especially effective for producing animations along story lines, where the characters must follow multiple objectives, while still accommodating geometric and kinematic constraints from the environment.

Abstract In this paper, we propose a new method to concatenate two dynamic full‐body motions such as punches, kicks, and flips by using the angular momentum as a cue. Through the observation of real humans, we have identified two patterns of angular momentum that make the transition of such motions efficient. Based on these observations, we propose a new method to concatenate two full‐body motions in a natural manner. Our method is useful for applications where dynamic, full‐body motions are required, such as 3D computer games and animations. Copyright © 2009 John Wiley & Sons, Ltd.

Efficient computation of strategic movements is essential to control virtual avatars intelligently in computer games and 3D virtual environments. Such a module is needed to control non-player characters (NPCs) to fight, play team sports or move through a mass crowd. Reinforcement learning is an approach to achieve real-time optimal control. However, the huge state space of human interactions makes it difficult to apply existing learning methods to control avatars when they have dense interactions with other characters. In this research, we propose a new methodology to efficiently plan the movements of an avatar interacting with another. We make use of the fact that the subspace of meaningful interactions is much smaller than the whole state space of two avatars. We efficiently collect samples by exploring the subspace where dense interactions between the avatars occur and favor samples that have high connectivity with the other samples. Using the collected samples, a finite state machine (FSM) called Interaction Graph is composed. At run-time, we compute the optimal action of each avatar by minmax search or dynamic programming on the Interaction Graph. The methodology is applicable to control NPCs in fighting and ball-sports games.

Abstract Evaluating the similarity of motions is useful for motion retrieval, motion blending, and performance analysis of dancers and athletes. Euclidean distance between corresponding joints has been widely adopted in measuring similarity of postures and hence motions. However, such a measure does not necessarily conform to the human perception of motion similarity. In this paper, we propose a new similarity measure based on machine learning techniques. We make use of the results of questionnaires from subjects answering whether arbitrary pairs of motions appear similar or not. Using the relative distance between the joints as the basic features, we train the system to compute the similarity of arbitrary pair of motions. Experimental results show that our method outperforms methods based on Euclidean distance between corresponding joints. Our method is applicable to content‐based motion retrieval of human motion for large‐scale database systems. It is also applicable to e‐Learning systems which automatically evaluates the performance of dancers and athletes by comparing the subjects' motions with those by experts. Copyright © 2008 John Wiley & Sons, Ltd.

We propose a data-driven approach to automatically generate a scene where tens to hundreds of characters densely interact with each other. During off-line processing, the close interactions between characters are precomputed by expanding a game tree, and these are stored as data structures called interaction patches. Then, during run-time, the system spatio-temporally concatenates the interaction patches to create scenes where a large number of characters closely interact with one another. Using our method, it is possible to automatically or interactively produce animations of crowds interacting with each other in a stylized way. The method can be used for a variety of applications including TV programs, advertisements and movies.

It is difficult to create scenes where multiple avatars are fighting / competing with each other. Manually creating the motions of avatars is time consuming due to the correlation of the movements between the avatars. Capturing the motions of multiple avatars is also difficult as it requires a huge amount of post-processing. In this paper, we propose a new method to generate a realistic scene of avatars densely interacting in a competitive environment. The motions of the avatars are considered to be captured individually, which will increase the easiness of obtaining the data. We propose a new algorithm called the temporal expansion approach which maps the continuous time action plan to a discrete space such that turn-based evaluation methods can be used. As a result, many mature algorithms in game such as the min-max search and α---β pruning can be applied. Using our method, avatars will plan their strategies taking into account the reaction of the opponent. Fighting scenes with multiple avatars are generated to demonstrate the effectiveness of our algorithm. The proposed method can also be applied to other kinds of continuous activities that require strategy planning such as sport games.

Skills to spin the ball are important for athletes in many sports such as baseball, soccer and tennis. Recently, researchers in aerodynamics and sports science started to analyze the correlation of the rotation and the trajectory of the ball; high speed cameras are used to analyze the rotation of the ball. However, the analysis had to be done either manually, or under special lighting condition. In this paper, we propose a new algorithm to track the translation and rotation of the ball shot by high speed cameras under outdoor ambient light condition. Using our system, the athletes can analyze their skills by shooting their motion and the resultant rotation of the ball. It can also be used as a tool to enhance sports TV programs by providing further scientific information to the audience.

We propose a new method to extract and calculate the 3D trajectory of a pitched baseball in a video clip. Compared to previous methods, which require video clips from multiple view points, only a single-view television clip is required for our method. Since global search methods based on dynamic programming are used to find the trajectory of the ball, the system is more robust than previous incremental methods. Therefore, our technique can be used to analyze the pitches not only in live TV programs, but also in previous games by famous pitchers. It is also possible to display the 3D trajectory of the baseball in a virtual environment from the viewpoint of the hitter. Pitchers can improve their skills by viewing the trajectory of their balls, and the hitters can view the pitches of various pitchers. As a result, our system can be used for baseball training, as well as for entertainment such as video games. The method to extract the ball from the scene is also applicable to sports such as tennis, volleyball, and soccer.