In this work, we introduce a novel high-fidelity full-head 3D avatar generation method from a single image, regardless of perspective, style, expression, or accessories. Prior works often fail to preserve consistent head geometry and facial details, primarily due to their limited capacity in modeling fine-grained facial textures and maintaining identity information. To address these challenges, we construct a new high-quality dataset containing 227 sequences of digital human portraits captured from 96 different perspectives, totalling 21,792 frames, featuring high-quality facial texture details. To further improve performance, we propose a novel multi-view diffusion named ID-TS diffusion model, which integrate identity and expression information into the two-stage multi-view diffusion process. The low-resolution stage ensures structural consistency of heads across multiple views, while the high-resolution stage preserves facial detail fidelity and coherence. Finally, we propose an enhanced feed-forward Gaussian avatar reconstruction method that optimizes the network on multi-view images of each single subject, significantly improving 3D facial texture details. Extensive experiments show that our method demonstrates robust performance across challenging scenarios, while showcasing broad applicability across numerous downstream tasks.

Novel view synthesis (NVS) from a single image is fundamental for multimedia applications including virtual/augmented reality and interactive 3D content retrieval. Despite notable progress, diffusion-based methods suffer from a persistent semantic gap: global textual conditioning lacks the granularity to preserve part-level attributes and inter-part spatial relationships under large viewpoint changes. We present Hierarchical Multi-Modal Novel View Synthesis (HM-NVS), a framework that bridges this gap through three synergistic contributions: (1) a hierarchical semantic extraction and adaptive fusion mechanism that integrates global, part-level, and geometric cues with view-dependent weighting; (2) a semantic-augmented 3D attention module that jointly enforces geometric and semantic consistency across generated views; and (3) an uncertainty-aware progressive refinement strategy that selectively allocates computation to difficult regions. Experiments on GSO, RTMV, and OmniObject3D demonstrate that HM-NVS improves PSNR by 15% and reduces LPIPS by 20% over state-of-the-art methods, with robust generalisation to non-rigid and articulated objects.

Video foundation models (VFMs) pretrained on internet-scale data have demonstrated remarkable generalization in semantic segmentation and object tracking. However, we identify a critical limitation when these models are deployed for quantitative industrial analysis: semantic understanding does not entail metrological precision. A VFM may correctly segment a rotating mechanical component while producing kinematic measurements that violate rigid-body constraints—a deficiency we term the semantic-precision gap. Through controlled experiments, we establish that this gap is structural: it persists under stronger backbones (SAM 3), parameter-efficient fine-tuning (LoRA, Adapters), and state-of-the-art video object segmentation (Cutie, DEVA), because none introduce the physical inductive biases necessary for geometric and kinematic reasoning.

Video Question-Answering (VideoQA) enables machines to interpret and respond to complex video content, advancing human-computer interaction. However, existing multimodal large language models (MLLMs) often provide incomplete or opaque explanations and existing benchmarks mainly focus on the correction of final answers, limiting insight into their reasoning processes and hindering both transparency and verifiability. To address this gap, we propose the Question Parsing, Video Alignment and Answer Aggregation framework (QPVA$^{3}$3), which leverages a compositional graph to drive visual and logical reasoning in VideoQA. Specifically, QPVA$^{3}$3 consists of three core components, the planner, executor, and reasoner to generate the compositional graph and conduct graph-driven reasoning. For the original question, the planner parses it into the compositional graph, capturing the underlying reasoning logic and structuring it into a series of interconnected questions. For each question in compositional graph, the executor aligns the video by selecting relevant video clips and generates answers, ensuring accurate, context-specific responses. For each question with its first-order descents, the reasoner aggregates answers by integrating reasoning logic with visual evidence, resolving conflicts to produce a coherent and accurate response. Moreover, to assess the performance of existing MLLMs in the reasoning processes of VideoQA, we introduce novel compositional consistency metrics and construct a VideoQA benchmark (QPVA$^{3}$3 Bench) with 3,492 question-video tuples, each annotated with detailed compositional graphs and fine-grained answers. We evaluate the QPVA$^{3}$3 framework on QPVA$^{3}$3 Bench and 5 other VideoQA benchmarks. Experimental results demonstrate that our framework improves both consistency and accuracy compared to baselines, leading to a more transparent and verifiable VideoQA system. This approach has the potential to advance the field, as supported by our comprehensive evaluation and benchmarking efforts.

Hyperspectral image classification is an attractive and challenging task due to the difficulty to acquire large labeled datasets, and its susceptibility to natural environmental influences. Deep learning-based methods can effectively enhance the efficiency of hyperspectral feature learning. Tensor-based approaches can effectively help preserving the inherent structural information. This paper introduces a novel tensor-based framework for hyperspectral image classification. It leverages the ‘bot-tleneck’ structure of an AutoEncoder to extract low-dimensional representations in a self-supervised manner. Additionally, local low-rank constraints in the embedding space facilitate the distribution of features across a low-rank manifold. Experiments conducted on real hyperspectral datasets demonstrate that the proposed method yields superior classification performance.

Abstract Currently, state-of-the-art object-tracking algorithms are facing a severe threat from adversarial attacks, which can significantly undermine their performance. In this research, we introduce MUNet, a novel defensive model designed for visual tracking. This model is capable of generating defensive images that can effectively counter attacks while maintaining a low computational overhead. To achieve this, we experiment with various configurations of MUNet models, finding that even a minimal three-layer setup significantly improves tracking robustness when the target tracker is under attack. Each model undergoes end-to-end training on randomly paired images, which include both clean and adversarial noise images. This training separately utilizes pixel-wise denoiser and feature-wise defender. Our proposed models significantly enhance tracking performance even when the target tracker is attacked or the target frame is clean. Additionally, MUNet can simultaneously share its parameters on both template and search regions. In experimental results, the proposed models successfully defend against top attackers on six benchmark datasets, including OTB100, LaSOT, UAV123, VOT2018, VOT2019, and GOT-10k. Performance results on all datasets show a significant improvement over all attackers, with a decline of less than 4.6% for every benchmark metric compared to the original tracker. Notably, our model demonstrates the ability to enhance tracking robustness in other blackbox trackers.

Semantic segmentation has achieved remarkable progress but remains challenging due to the complex scene, object occlusion, and so on. Some research works have attempted to use extra information such as a depth map to help RGB based semantic segmentation because the depth map could provide complementary geometric cues. However, due to the inaccessibility of depth sensors, depth information is usually unavailable for the test images. In this paper, we leverage only the depth of training images as the privileged information to mine the hard pixels in semantic segmentation, in which depth information is only available for training images but not available for test images. Specifically, we propose a novel Loss Weight Module, which outputs a loss weight map by employing two depth-related measurements of hard pixels: Depth Prediction Error and Depth-aware Segmentation Error. The loss weight map is then applied to segmentation loss, with the goal of learning a more robust model by paying more attention to the hard pixels. Besides, we also explore a curriculum learning strategy based on the loss weight map. Meanwhile, to fully mine the hard pixels on different scales, we apply our loss weight module to multi-scale side outputs. Our hard pixels mining method achieves the state-of-the-art results on three benchmark datasets, and even outperforms the methods which need depth input during testing.

Video data are of two different intrinsic modes, in‐frame and temporal. It is beneficial to incorporate static in‐frame features to acquire dynamic features for video applications. However, some existing methods such as recurrent neural networks do not have a good performance, and some other such as 3D convolutional neural networks (CNNs) are both memory consuming and time consuming. This study proposes an effective framework that takes the advantage of deep learning on the static image feature extraction to tackle the video data. After extracting in‐frame feature vectors using a pretrained deep network, the authors integrate them and form a multi‐mode feature matrix, which preserves the multi‐mode structure and high‐level representation. They propose two models for follow‐up classification. The authors first introduce a temporal CNN, which directly feeds the multi‐mode feature matrix into a CNN. However, they show that characteristics of the multi‐mode features differ significantly in distinct modes. The authors therefore further propose the multi‐mode neural network (MMNN), in which different modes deploy different types of layers. They evaluate their algorithm with the task of human action recognition. The experimental results show that the MMNN achieves a much better performance than the existing long short‐term memory‐based methods and consumes far fewer resources than the existing 3D end‐to‐end models.

Rich clothes datasets and high-quality annotations have driven recent advances in fashion clothes recognition. However, the existing approaches treat clothes as common images, ignoring the prior clothing knowledge such as spatial relations, symmetry, proportions, and key characteristics of clothes. In order to combine the semantic information with the advantages of deep learning, we propose Detection+, a model using the prior symmetric constraint to refine the keypoints located by any backbone detection networks. To deal with uncertainty in labelling clothing, we introduce a new loss to utilize all available data which contain "maybe" labels. Detection+ has reduced about 2.54% Normalized Error in FashionAI dataset and improved 3.2% AP in human keypoints dataset coco2017 compared to the Mask R-CNN baseline. A large number of experimental results show the proposed approach achieves better results in different recognition datasets (resp., FashionAI, and Deepfashion) with about (resp., 2.57% mAP, and 10% recall) improvements.