Training-efficient adaptation of video action recognition models remains challenging because high-performing architectures often require costly end-to-end retraining and dense optical-flow pre-processing. This paper presents MoCLIP, a modular framework that combines a frozen Contrastive Language–Image Pre-training (CLIP) image encoder with a lightweight motion-vector (MV) pathway operating directly in the compressed video domain. The appearance branch provides semantic image features, while the motion branch extracts temporal cues from H.264 motion vectors through a temporal-segment-style encoder. A cross-modal feature integration (CMFI) module fuses these complementary representations, and the high-compute variant applies fixed arithmetic probability fusion (PF) over multiple prediction branches. By keeping the semantic backbone fixed and training only the compact motion encoder and fusion heads, MoCLIP-XL has 13.53 million trainable parameters under the UCF101 reference configuration used for resource accounting and requires no backpropagation through CLIP. The XL variants reach 96.5 https://github.com/microa/MoCLIP .
Sparse-view neural radiance field reconstruction remains challenging because limited observations provide insufficient constraints for recovering reliable scene geometry and appearance. This paper presents ConRF, a sparse-view neural rendering framework built upon a generator-parameterized factorized radiance field. ConRF introduces image-conditioned tensor generation and rendered-depth-guided trajectory consistency to incorporate observation-level information into both representation construction and scene optimization. Specifically, global descriptors extracted from the available input views are aggregated into a scene-level condition and injected into the latent representation and tensor generator through learnable projections. In parallel, rendered depth is used to establish pixel–3D–pixel correspondences between calibrated input views, providing additional cross-view photometric constraints during optimization. Reliability filtering and a warm-up strategy are adopted to reduce the influence of unstable correspondences at early training stages. We evaluate the proposed method on NeRF-Synthetic, OpenIllumination, and Tanks and Temples under sparse-view settings. The results show that ConRF improves the overall reconstruction quality of the underlying baseline across the evaluated benchmarks, with more evident gains when the available observations provide sufficient cross-view constraints. Additional ablation, sensitivity, and pose-perturbation experiments further analyze the contributions and behavior of the proposed components.
Person re-identification in unconstrained surveillance remains challenging because identity cues are frequently disrupted by occlusion, viewpoint variation, background clutter and redundant attention in multi-branch feature learners. This study proposes an adaptive multi-granularity aggregation framework based on a vision state space model to improve robust and efficient visual retrieval without depending on external pose, segmentation or vision-language priors at inference. Integrating state-space visual computing into the framework, the model first uses two-dimensional cross scanning to model long-range spatial dependencies with linear complexity and to produce global and local token representations. It then estimates discriminative visible regions through cumulative energy distribution of activation responses, replacing fixed body partitioning with image-specific occlusion-robust recognition. To reduce branch homogenisation, a token-competition mechanism combines inner saliency focusing with outer semantic repulsion, encouraging different local branches to capture complementary identity evidence such as clothing texture, accessories and shoe details. A context-aware gated aggregation module further evaluates global–local consistency and suppresses occlusion-contaminated local responses, while confidence-driven routing switches between fast global retrieval and fine-grained local processing for difficult samples. Experiments are reported on Market-1501, MSMT17, Occluded-Duke and Occluded-REID using Rank-k and mAP metrics. The method achieves 96.8
High-resolution large-format scanned core images contain rich microstructural information, but pronounced scale variations and heterogeneous component distributions make pixel-level segmentation challenging. Conventional convolutional neural networks capture local details effectively but are limited in modeling long-range spatial dependencies. Global modeling structures can improve contextual representation, but they often underuse local and multi-scale information. To address these issues, we propose a dual-encoder feature fusion network for semantic segmentation of core images. The network combines a residual convolutional branch for local representation with a visual state-space branch for global contextual modeling. A Multi-Scale Adaptive Fusion Module dynamically integrates local and global features. A Multi-Pooling Channel–Spatial Feature Refinement Module further enhances fused skip-connection features along the channel and spatial dimensions. Experiments show that the proposed method achieves mIoU, mPrecision, mRecall, mF1, and mAcc values of 87.24
The use of facial expression recognition has received considerable attention since it has been applied in human–computer interaction, healthcare, surveillance, and affective computing. This paper describes all the steps in the classification and feature extraction methodology applied to the task of recognizing emotions based on face images and video sequences. The conventional machine learning methods, which are K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Tree, and Random Forest, are trained using handcrafted features, such as geometric distances, Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), and facial action units. These approaches perform well in controlled environments but tend to be weak in complex, real-world scenarios. The most recent development in deep learning, specifically Convolutional Neural Networks (CNN), has allowed automatic feature extraction and end-to-end learning, which have greatly increased the recognition accuracy. Modern architectures like ResNet, VGG, MobileNet, and ensemble CNN models are highly effective when applied to benchmark datasets. Moreover, hybrid models that incorporate CNN with other classifiers such as SVM and Random Forest increase the efficiency of classifications. In the case of video-based emotion recognition, Recurrent Neural Networks (RNN), such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks, are used to identify temporal dependencies and enhance results by modeling the dynamics of facial expressions. The CNN-RNN hybrid frameworks and spatio-temporal networks further enhance spatial and temporal feature learning. Despite these developments, there are still a number of issues, such as weak generalization in non-constrained settings, excessive computational demands, the need to use large labeled sets, and the inability to detect subtle and compound emotions. This review focuses on the important gaps in research and future directions like lightweight model design, multimodal integration, self-supervised learning, and advanced temporal models. Unlike prior narrative surveys in this area, this review follows a PRISMA-documented systematic search and screening protocol, explicitly positions its coverage against comparable existing surveys, and synthesizes quantitative, metric-, and validation-protocol-matched comparisons across studies rather than relying on accuracy figures alone. The results will provide clear insight into the existing methodologies and inform the construction of effective and powerful facial expression recognition systems.
