In the mission of 2D-to-3D lifting for human pose estimation, the self-attention mechanism has demonstrated remarkable efficacy in capturing global interdependencies among joints. However, the attention matrix generated by the Query and Key, derived from the mapping of joint features, exclusively captures implicit spatial relationships among these features. In this paper, we propose that integrating the physical topological relationships of the human body and the explicit spatial relationships among the joints can enhance the overall representational capacity. Therefore, we introduce an Adaptive Multi-Relationship Attention (AMuRA), which adaptively captures diverse global relationships among human body joints by incorporating the joint Self-Similarity Matrix (SSM) and predefined skeleton adjacency matrix into the traditional attention mechanism. Additionally, graph convolutional networks (GCNs) can capture local representations between adjacent joints. By combining AMuRA with the GCN Block, we propose a Multi-Relationship Encoder (MuRE) that captures multi-layer dependencies among human body joints through a global-to-local architecture.

Transformer-based 3D human pose estimation (HPE) methods face efficiency-accuracy trade-offs due to self-attention’s quadratic complexity. While State Space Models (SSMs) offer linear complexity and strong long-range modeling, direct Mamba adaptation to video-based 3D HPE is suboptimal, as unidirectional SSMs poorly capture spatio-temporal context. To address this, we propose DBMambaPose, a novel attention-free 3D HPE architecture based on SSMs. Its core DBMambaPose Block alternately stacks Spatial Disentangled Bidirectional Mamba Block (S-DBMB) for intra-frame joint spatial dependencies and Temporal Disentangled Bidirectional Mamba Block (T-DBMB) for inter-frame joint motion trajectories. We further introduce a Decoupled Spatial-Temporal Bidirectional Scanning mechanism (DST-BS) to enable frame-ordered spatial bidirectional processing and joint-ordered temporal bidirectional processing, enhancing fine-grained spatio-temporal feature modeling. Four DBMambaPose variants provide flexible accuracy-efficiency trade-offs. Experiments on Human3.6M and MPI-INF-3DHP show DBMambaPose achieves state-of-the-art performance with reduced computational costs.

Most existing methods for 3D human pose estimation from monocular images focus on learning the spatial correlation of either the global or local joints of the human body but fail to adequately capture the inherent dependencies between them. To address this limitation, we propose the Interactive Complementary Fusion Network (ICFNet), an algorithm designed to fully utilize the prior knowledge of both global and local joint relationships to enhance prediction performance. Specifically, we introduce two feature capturers: the Global Knowledge Prior Capturer (GKPC) and the Local Region Subject Capturer (LRSC), which respectively capture global body knowledge and local joint information. Additionally, we propose three joint constraint mechanisms to express the potential association dependencies between global and local joints, which are further modeled using two association capturers: the Refined-Regression Association Capture Module (RR-ACM) and the Generalized-Guidance Association Capture Module (GG-ACM). Moreover, we introduce a novel feature transformation module, the Link Conversion Module (LCM), to transform and augment pose features. The algorithm adopts a complementary process to enhance the interaction and fusion of global and local feature information by gradually imposing constraints on the physical topological features of the human body, thereby improving its modeling capabilities. Extensive experiments demonstrate that our proposed ICFNet achieves state-of-the-art results on two challenging benchmark datasets: Human 3.6M and MPI-INF-3DHP.

Low-light images often have defects such as low visibility, low contrast, high noise, and high color distortion compared with well-exposed images. If the low-light region of an image is enhanced directly, the noise will inevitably blur the whole image. Besides, according to the retina-and-cortex (retinex) theory of color vision, the reflectivity of different image regions may differ, limiting the enhancement performance of applying uniform operations to the entire image. Therefore, we design a Hierarchical Flow Learning (HFL) framework, which consists of a Hierarchical Image Network (HIN) and a normalized invertible Flow Learning Network (FLN). HIN can extract hierarchical structural features from low-light images, while FLN maps the distribution of normally exposed images to a Gaussian distribution using the learned hierarchical features of low-light images. In subsequent testing, the reversibility of FLN allows inferring and obtaining enhanced low-light images. Specifically, the HIN extracts as much image information as possible from three scales, local, regional, and global, using a Triple-branch Hierarchical Fusion Module (THFM) and a Dual-Dconv Cross Fusion Module (DCFM). The THFM aggregates regional and global features to enhance the overall brightness and quality of low-light images by perceiving and extracting more structure information, whereas the DCFM uses the properties of the activation function and local features to enhance images at the pixel-level to reduce noise and improve contrast. In addition, in this paper, the model was trained using a negative log-likelihood loss function. Qualitative and quantitative experimental results demonstrate that our HFL can better handle many quality degradation types in low-light images compared with state-of-the-art solutions. The HFL model enhances low-light images with better visibility, less noise, and improved contrast, suitable for practical scenarios such as autonomous driving, medical imaging, and nighttime surveillance. Outperforming them by PSNR = 27.26 dB, SSIM = 0.93, and LPIPS = 0.10 on benchmark dataset LOL-v1.